Abstract: This paper develops and explores a first-order theory of time that appears useful as an underlying framework for a wide range of practical applications in artificial intelligence (AI). In particular, it presents a concise, formal axiomatization of “interval-based” time as described by Allen and then explores the relationship between interval-based and point-based temporal theories in detail. This analysis should be useful to both theoretical and practical work in AI that involves the representation of time, since it shows what distinctions are actually substantive and what arise merely from formalisms being notational variants of one another. Cet article elabore et explore une theorie de premier ordre du temps qui semble utile comme cadre de travail dans une vaste gamme d'applications pratiques dans le domaine de l'intelligence artificielle. En particulier, il presente une axiomatisation concise du temps a base d'intervalles telle que decrite par Allen; de plus, il explore la relation entre les theories temporelles a base d'intervalles et celles a base de points. Cette analyse devrait ětre utile dans les travaux pratiques et theoriques qui necessitent la representation du temps, car elle demontre quelles distinctions sont vraiment importantes et lesquelles decoulent de formalismes qui sont chacun des variantes notationnelles de l'autre.

Abstract: A hierarchical planning system achieves efficiency by planning with the most important conditions first, and considering details later in the planning process. Few attempts have been made to formalize the structure of the planning knowledge for hierarchical planning. For a given domain, there is usually more than one way to define its planning knowledge. Some of the definitions can lead to efficient planning, while others may not. In this paper, we provide a set of restrictions which defines the relationships between a non-primitive action and its set of subactions. When satisfied, these restrictions guarantee improved efficiency for hierarchical planning. One important feature of these restrictions is that they are syntactic and therefore do not depend on the particular structure of any plan. Along with these restrictions, we also provide algorithms for preprocessing the planning knowledge of a hierarchical planner. When used during planning, the preprocessed operator hierarchies can enable a planner to significantly reduce its search space. Un systeme de planification hierarchique permet d'obtenir un rendement efficace dans la mesure ou il planifie d'abord les conditions importantes avant de s'attarder aux details secondares dans le processus de planification. Quelques ten-tatives ont ete realisees en vue de formaliser la structure des connaissances de planification dans la planification hierarchique. Dans un domaine donne, il existe normalement plus d'une facon de definir les connaissances de planification. Certaines definitions peuvent permettre une planification efficace et d'autres non. Dans cet article, l'auteur propose une serie de restrictions qui definissent les rapports entre une action non primitive et son ensemble de sous-actions. Lorsqu'elles sont respectees, ces restrictions permettent d'ameliorer l'efficacite dans la planification hierarchique. l'une des caracteristiques importantes de ces restrictions est qu'elles sont syntaxiques et ne dependent done pas de la structure particuliere d'un plan. En plus de ces restrictions, l'auteur propose egalement des algorithmes pour le pretraitement des connaissances de planification d'un planificateur hierarchique. Quand elles sont utilisees durant la planification, les hierarchies d'operateur peuvent permettre au planificateur de reduire considerablement son espace de recherche.

Abstract: This article ties together previously scattered research on discontinuous grammars—logic grammars in which non-explicit sequences of symbols can be alluded to in the rules and sometimes repositioned by them. After an introduction, we define them formally, present their background, and provide intuitive insight into their use. Next, we examine several motivating arguments, from both formal and natural language processing viewpoints, and we discuss the static discontinuity family of these grammars, in which (a) the nonexplicit strings are not allowed to move and (b) linguistic constraints specifically designed to suit, in particular, Government and Binding theory can be defined modularly and statically in terms of node domination in parse trees and are enforced dynamically. Finally, we discuss implementation issues, related work, and extensions. Cet article presente une synthese des recherches anterieures sur les grammaires discontinues, c'est-a-diregles grammaires logiques dans lesquelles des sequences de symboles non explicites peuvent faire l'objet d'un renvoi dans les regles et peuvent, parfois, ětre repositionnees par celles-ci. Apres une introduction, nous les definissons de facon formelle, presentons leur historique et fournissons une explication sur leur utilisation. Ensuite, nous examinons plusieurs arguments motivants des points de vue du traitement naturel et formel du langage, et discutons egalement de la famille de discontinuite statique de ces grammaires, dans laquelle (a) les chaines non explicites n'ont pas le droit de bouger et (b) les contraintes linguistiques concues specialement pour ětre conformes a la theorie gouvernement-liage peuvent ětre definies de facon modulaire et statique en termes de domination des noeuds dans les arbres d'analyse, et mises en application de maniere dynamique. Enfin, les questions d'implementation, les travaux connexes et les extensions sont discutes.

Abstract: The use of a single grammar in natural language parsing and generation is most desirable for a variety of reasons, including efficiency, perspicuity, integrity, robustness, and a certain amount of elegance. These characteristics have been noted before by several researchers, but it was only recently that more serious attention started to be paid to the problem of creating a bidirectional system for natural language processing. In this paper we discuss a somewhat more radical version of the problem: given a parser for a language, can we reverse it so that it becomes an efficient generator for the same language? Furthermore, since both the parser and the generator are based upon the same grammar, are there any normalization conditions upon the form of the grammar that must be met in order to assure the maximum efficiency of the reversed program? Can other grammars be transformed into the normal form? We describe the results of an experiment with PROLOG-based logic grammar which has been derived from a substantial-coverage string grammar for English. We present an alogorithm for automated inversion of a unification parser into an efficient unification generator, using the collections of minimal sets of essential arguments for predicates. We discuss the scope of the present version of the algorithm and then point out several possible avenues for extension. We also outline a preliminary solution to the question of grammar's “normal form” and suggest a handful of normalizing transformations that can be used to enhance the efficiency of the generator. This research interacts closely with a Japanese-English machine translation project at New York University, for which the first implementation of the inversion algorithm has been prepared. Ľutilisation ?une grammaire unique dans ľanalyse et la generation du langage naturel est tres souhaitable pour une multitude de raisons, dont ľefficacite, la clarte, ľintegrite, la solidite et une certaine dose ?elegance. Ces caracteristi-ques ont deja ete constatees par plusieurs chercheurs; cependant, ce n'est que tout recemment qu'une attention plus serieuse a ete portee au probleme de la creation ?un systeme bidirectionnel de traitement du langage naturel. Dans cet article, ľauteur discute ?un aspect un peu plus radical du probleme qui concerne la possibilitye?inverser le parseur ?un langage afin qu'il devienne un generateur efficace dans le měme langage. En outre, comme le parseur et le generateur sont fonction de la měme grammaire, existe-t-il des conditions de normalisation de la forme de grammaire qu'il faut respecter afin de s'assurer de ľefficacite du programme inverse? Y a-t-il ?autres gramaires qui peuvent ětre transformers en forme normale?Ľauteur decrit les resultats ?une experience avec une grammaire logique basee sur PROLOG qui est derivee ?une grammaire en chaines de langue anglaise. II propose un algorithme pour ľinversion automatique ?un parseur ?unification en generateur ?unification efficace, aľaide des ensembles minimaux ?arguments essentiels pour predicats. II discute de la portee de la version actuelle de ľalgorithme et souligne diverses possibilityes de developpement. II elabore egalement une solution preliminaire a la question de la «forme normale» de la grammaire et suggere un certain nombre de transformations qui peuvent servir a accroitre ľefficacite du generateur. Cette recherche est realisee en etroite collaboration avec un projet de traduction-machine du japonais vers ľanglais pour lequel la premiere mise en oeuvre de ľalgorithme ?inversion a ete realisee. Mots cles: traitement du langage naturel, analyse, generation, grammaire logique, grammaire reversible, arguments essentiels, predicat inverse, direction des calculs.

Abstract: Human discourse is fraught with rhetorical devices such as contradictions, illustrations, and analogies. These rhetorical devices carry important information which a listener uses to speed up the comprehension process. In this paper, we present a taxonomy of rhetorical devices commonly used in tutoring environment, and model the meaning of a class of rhetorical devices in terms of their anticipated effect on a listener's knowledge. These predictions are then used in planning computer-generated discourse. As a testbed for our ideas, a system called WISHFUL is being implemented to generate commentaries in the domain of high-school algebra within the framework of an intelligent tutoring system.

Abstract: Cartographic map overlay is the process of superimposing two maps into one to convey information in spatial correlation. A map refers to one in vector representation: a two-dimensional spatial data structure of nodes, chains, and polygons. We present a map overlay system developed in Prolog. The system adopts a relational approach to data structuring. We represent geometric entities and their relationships as facts, and encode geometry algorithms in the rules. Set-based operations perform data processing. To speed up the search for chain intersections, a uniform rectangular grid is imposed over the object space for spatial sorting by distribution. We sort out potentially intersecting edge segments to those occupying some common grid cell. Each bucket, if non-empty, is implemented as a Prolog fact identifying the grid cell for random access. Geometric intersections are calculated using exact rational arithmetic implemented in Prolog. Numerical accuracy is preserved and we can identify all the special cases of tangent conditions. We can then guarantee topological consistency, and stability in the process of map overlay is therefore achieved. La confection de cartes geographiques necessite la superposition de deux cartes afin que Ton puisse mettre en cor-respondance des informations sur le plan spatial. La reference ?une carte a une autre se fait sous forme de representation vectorielle, c'est-a-dire une structure bidimensionnelle de donnees spatiales de noeuds, de chaines et de polygones. Cet article presente un systeme de superposition de cartes elabore dans Prolog. Le systeme favorise une approche rela-tionnelle 4aG la structuration des donnees. Des entites geometriquss sont representees ainsi que leur relation comme faits; des algorithmes de geometrie sont ensuite encodes dans les r4eGgies. Des operations basees sur des ensembles effectuent le traitement des donnees. Afin ?accelerer la recherche des intersections de chaines, une grille rectangulaire uniforme est superposee sur ľespace de ľobjet pour permettre un tri spatial par repartition. On distingue les segments de contours qui font ľobjet ?intersections de ceux qui occupent une cellule commune de la grille. Chaque cuvette qui n'est pas vide est traitee comme un fait Prolog identifiant la cellule en vue ?un acces aleatoire. Les intersections geometriques sont calculees aľaide ?une arithmetique rationnelle exacte incorporee a Prolog. La precision numerique est preservee et il est possible ?identifier tous les cas speciaux de conditions tangentes. Nous pouvons ensuite garantir la consistance topologique et assurer la stabilitye du processus de superposition de cartes.

Abstract: This paper presents a description of the application of an attributed graph based approach to the synthesis of a geometric world model for use in navigation by a mobile robot. Our aim is to develop the theoretical aspects of graph synthesis for mobile robot world knowledge acquisition, and to demonstrate the validity of the approach with a simulation before implementation on the rover. A boundary representation of free space consisting of directed line segments organized into a directed attributed graph is used. The synthesis problem can be considered as having two parts: matching of a local model with a global model and the construction of a new global model. Structural and geometric local and global constraints are used to limit and direct the search for valid graph mappings. The constraints are the source of rules for matching primitives and graphs and are used in the process of constructing a new world model graph. An algorithm for graph synthesis is implemented in a software simulation for testing and experimentation. Cet article presente une description de ľapplication ?une methode basee sur les graphes attribues a la synthese ?un modele geometrique du monde utilise par un robot mobile en navigation. Le but est de developper les aspects theori-ques de la synthese des graphes pour ľacquisition de connaissances du monde par le robot et de demontrer la validite de la methode par une simulation avant la mise en oeuvre. Une representation des frontieres de l'espace libre formees de segments de lignes orientes, organises sous forme de graphe attribue orient4eA, est utilisee. On peut considerer le probleme de synthese comme ayant deux parties: la mise en relation ?un modele local avec un modele global et la construction ?un nouveau modele global. Les contraintes locales et globales sur les plans de la structure et de la geometrie sont utilisees pour limiter et orienter la recherche de mises en relation valides de graphes. Les contraintes sont la source de regies pour le couplage de graphes et de primitifs et elles sont utilisees pour construire un nouveau graphe de modele du monde. Un algorithme pour la synthese de graphe est mis en oeuvre par le biais ?une simulation afin de le soumettre a des essais et des experiences.

Abstract: McAllester's conspiracy numbers algorithm is an exciting new approach to minimax search that builds trees to variable depth without application-dependent knowledge. The likelihood of the root taking on a value is expressed by its conspiracy number: the minimum number of leaf nodes that must change their value to cause the root to change to that value. This paper describes experiences with the algorithm, using both random and application-generated trees. Experiments suggest that more emphasis on breadth, rather than depth, can lead to significant performance improvements. New enhancements to the algorithm are capable of solving 41% more problems than McAllester's original proposal. L' algorithme des nombres de la conspiration de McAllester est une toute nouvelle technique de recherche minimax qui permet de construire des arbres de profondeur variable sans avoir recours aux connaissances basees sur les applications. La probabilitye qu'une racine prenne une valeur est exprimee par son nombre de conspiration, soit le nombre minimum de noeuds feuilles qui doivent changer de valeur afin que la racine prenne cette valeur. Cet article decrit les experiences avec l'algorithme a aide d'arbres generes par les applications et d'arbres aleatoires. Ces experiences suggerent qu'on peut obtenir des ameliorations importances de la performance en accordant plus d'importance a la largeur qu'a la profondeur. Des ameliorations de l'algorithme permettent d'augmenter le pourcentage de resolution des problemes de 41%, en comparaison avec la technique originale de McAllester.

Abstract: In this paper, we present a formalism called feature grammar and its application to several problems of semantic analysis. Our extension concerns the structure of the feature value sets, which can be complex, and the definition of unification, which is dependent on this structure. Moreover, we introduce generation rules for feature symbols in order to determine well-formed symbols, which form the alphabet of a formal language for natural language analysis. Dans cet article, ľauteur traite ?un formalisme appele grammaire de caracteristiques et de son application a la resolution de plusieurs problemes ?analyse semantique. Ce developpement touche a la structure des ensembles de valeurs des caracteristiques qui peuvent ětre complexes et a la definition de ľunification qui est fonction de cette structure. De plus, des regies de generation sont proposees pour les symboles des caracteristiques afin de determiner des symboles bien formes qui constituent ľalphabet ?un langage formel pour ľanalyse du langage naturel. Mots cles: traitement du langage naturel, grammaire de caracteristiques, analyse semantique.

Abstract: In this paper, we present an extension to PROLOG we call DISLOG which is designed to deal with relations between non-contiguous elements in a structure. This extension turns out to be well suited for syntactic analysis of natural and artificial languages. It is also well adapted to express traversal constraints in applications such as planning and expert systems and deductive systems involving, for example, temporal reasoning, DISLOG belongs to the constrained logic programming paradigm and turns out to be more declarative, transparent, and simple than PROLOG to deal with longdistance relations. Dans ce texte, nous presentons une extension a PROLOG, appelee DISLOG qui permet de traiter des phenomenes de dependance a longue distance dans un arbre de preuve. Cette extension apparait ětre bien adaptee aľanalyse syntaxique des langages naturels et artificiets. Elle est aussi bien adaptee aľexpression de contraintes transversales dans les applications telles que la planification. DISLOG, qui appartient au paradigme des programmes logiques contraints, est plus simple, plus transparent et plus declaratif que PROLOG pour traiter des relations a longue distance.

Abstract: There are two main issues I want to address in this response to Peter Cheeseman’s “An inquiry into computer understanding” (hereafter, abbreviated as Inquiry) and “In defense of An inquiry into computer understanding” (hereafter, abbreviated as Defense): (1) the “right” approach to reasoning about uncertainty and (2) the relationship between logic and probability. I Cheeseman’s positions on these issues is at times confusing, and occasionally contradictory (as he admits in Defense). Rather than belaboring the contradictions, I will try to focus on the high-level issues, which I believe are important ones for AI.

Abstract: This paper addresses the problem in inductive generalization of determining when a general hypothesis is supported by a particular instance. If we accept that, first, some facts do indeed support a general hypothesis and, second, that an instance that supports a hypothesis also supports all logical consequences of the hypothesis, then unintuitive and problematic results are immediately forthcoming. These assumptions lead, for example, to the conclusion that a blue Honda is confirming evidence for the hypothesis that ravens are black. This problem is variously known as the paradoxes of confirmation or Hempel's paradox. In this paper I develop a formal characterization of the problem. The assumption that whatever supports all classical consequences of the hypothesis is rejected. Rather, I argue that a weaker notion of consequence should be adopted for determining what consequences of a hypothesis are supported by the same evidence. An extant formal system for learning from examples is used to address these problems of evidential support, and it is shown that in this framework the problems do not arise. Cet article traite du probleme qui se pose lorsqu'il s'agit de determiner quand une hypothese generale est corroboree par une instance. Si nous tenons pour acquis que certains faits corroborent une hypothese generale et qu'une instance qui appuie une hypothese corrobore egalement toutes les consequences logiques de cette hypothese, alors les resultats non intuitifs et problematiques se produisent immediatement. Par exemple, ces hypotheses conduisent a la conclusion qu'une voiture Honda bleue est l'evidence que les corbeaux sont noirs. Ce probleme est souvent appele le paradoxe de la confirmation ou le paradoxe d'Hempel. Dans cet article, l'auteur elabore une caracterisation formelle du probleme. L'hypothese a l'effet que tout ce qui appuie une hypothese corrobore egalement toutes les consequences classiques de l'hypothese est rejetee. L'auteur soutient plutot qu'une notion plus faible de la consequence devrait ětre adoptee pour determiner les consequences d'une hypothese qui sont corroborees par la měme evidence. Un systeme formel d'apprentissage a partir des exemples est utilise pour solutionner ces problemes d'evidence et il est demontre que dans ce cadre de travail, les problemes ne se posent pas.

Abstract: A synergism has begun to surface from the artificial intelligence (AI) and engineering communities: an effort to apply AI techniques to engineering problem-solving activities, and to study problems arisen from various engineering fields as a way to develop AI theories and methodologies. The paper first discusses the needs of such a synergical approach and identifies in a broad perspective some AI techniques currently being applied to engineering. It then describes a system, called KREATOR, which applies qualitative reasoning, a subfield of AI, to computer-aided design (CAD). KREATOR is a knowledge capturing scheme that allows the designers to record their qualitative knowledge of how mechanical devices behave. KREATOR then automatically generates qualitative simulations

Abstract: I think the picture of “Robbie” in Cheeseman’s rejoinder is one in which few people would find much to quarrel with. However, 1 was a little disappointed in seeing Cheeseman “stand firm” on some claims that just don’t stand up to scrutiny. In particular, he still implies that truth and falsify preclude degrees of belief. I cannot fathom why anyone would think it impossible for a proposition to be bofh true and believed to degree 0.8, or bofh false and believed to degree 0.5, and so on. He also says that invoking the notion of truth in analyzing a method of belief computation is “irrelevant to the logicist program.” That amounts to saying that logicists are not concerned with the semantic underpinnings of inference procedures, a patent falsehood. His argument against the practical relevance of truth values, because they are “unobtainable quantities,” has two flaws. First, they are in some cases obtainable, namely in the case of valid formulas and logical falsehoods (and the former include most of the mathematics!), and perhaps in the case of primal sensory data (as Cheeseman himself assumes). Second, even in physics which I suppose is his model here “unobtainable quantities” are of crucial importance. For example, the position of an electron cannot be determined precisely (it would have to have infinite momentum); in fact, it can’t even be bounded with absolute certainty, because any measurement we take may be an artifact of quantum-mechanical perturbations of the measuring apparatus. We have to be satisfied with getting “very likely” bounds for the position of the electron. Yet, all of quantum mechanics is couched in the language of these strictly “unobtainable” quantities, such as position, momentum, energy, etc. The same goes for truth values we can often be pructically certain of the truth values of various propositions, but, in principle, they are usually strictly “unobtainable.” Finally, the reply to my “ace of spades” example completely misses the point. Cheeseman claims that universal instantiation is disallowed in this instance. But the very meaning of universal quantification, if it is to have its usual meaning, is that it holds for all possible values of the variable! If universal instantiation is disallowed, then all I can say is that Cheeseman is using V to mean something entirely different from universal quantification, which is confusing at best. Furthermore, his attempt to escape contradiction would be right if x were a random variable, which he denies. It certainly isn’t right as it stands, and the error does not lie with Cox, or de Finetti, or Carnap who all understood universal quantification, and incidentally thought logic and probability entirely compatible.

Abstract: I think the picture of “Robbie” in Cheeseman’s rejoinder is one in which few people would find much to quarrel with. However, 1 was a little disappointed in seeing Cheeseman “stand firm” on some claims that just don’t stand up to scrutiny. In particular, he still implies that truth and falsify preclude degrees of belief. I cannot fathom why anyone would think it impossible for a proposition to be bofh true and believed to degree 0.8, or bofh false and believed to degree 0.5, and so on. He also says that invoking the notion of truth in analyzing a method of belief computation is “irrelevant to the logicist program.” That amounts to saying that logicists are not concerned with the semantic underpinnings of inference procedures, a patent falsehood. His argument against the practical relevance of truth values, because they are “unobtainable quantities,” has two flaws. First, they are in some cases obtainable, namely in the case of valid formulas and logical falsehoods (and the former include most of the mathematics!), and perhaps in the case of primal sensory data (as Cheeseman himself assumes). Second, even in physics which I suppose is his model here “unobtainable quantities” are of crucial importance. For example, the position of an electron cannot be determined precisely (it would have to have infinite momentum); in fact, it can’t even be bounded with absolute certainty, because any measurement we take may be an artifact of quantum-mechanical perturbations of the measuring apparatus. We have to be satisfied with getting “very likely” bounds for the position of the electron. Yet, all of quantum mechanics is couched in the language of these strictly “unobtainable” quantities, such as position, momentum, energy, etc. The same goes for truth values we can often be pructically certain of the truth values of various propositions, but, in principle, they are usually strictly “unobtainable.” Finally, the reply to my “ace of spades” example completely misses the point. Cheeseman claims that universal instantiation is disallowed in this instance. But the very meaning of universal quantification, if it is to have its usual meaning, is that it holds for all possible values of the variable! If universal instantiation is disallowed, then all I can say is that Cheeseman is using V to mean something entirely different from universal quantification, which is confusing at best. Furthermore, his attempt to escape contradiction would be right if x were a random variable, which he denies. It certainly isn’t right as it stands, and the error does not lie with Cox, or de Finetti, or Carnap who all understood universal quantification, and incidentally thought logic and probability entirely compatible.

Abstract: Halperri argues for alternative non-Bayesian approaches to uncertainty based on problems he perceives in the Bayesian approach. In particular, he argues for a distinction between degrees of belief and statistical statements (based on the concept of random sampling). In this response I show that there is no difference between these two concepts in the Bayesian framework, and that the replacement of variables by constants in probabilistic predicate calculus expressions is valid, despite Halpern's objections. The main reason for his rejection of the simpler approach is that he does not condition his belief statements on the evidence used to form these beliefs, and so gets into trouble when new evidence is received. This failure to properly take evidence into account invalidates most of his other criticisms. While I approve of his call for more formal rigor in representing Bayesian practice, his claim to have provided a semantics is misleading – what he has provided is not operational.

Abstract: Allen gives an algebra for representing qualitative temporal information about the relationships between pairs of intervals. In this paper, we address a fundamental reasoning task that arises in applications of the algebra: Given (possibly indefinite) knowledge about the relationships between intervals, find all feasible relationships between two intervals. We call this the minimal labels problem. Finding the minimal labels can be viewed as computing the deductive consequences of our knowledge. Determining exact solutions to this problem has been shown to be (almost assuredly) intractable. Allen gives an approximation algorithm based on constraint propagation. We present new approximation algorithms, determine analytically under what conditions the algorithms are exact, and examine, through some computational experiments, the quality of the approximate solutions produced by the algorithms. We also give a simple test for predicting when the approximation algorithms will and will not produce good quality approximations. Finally, we survey three example applications of the interval algebra chosen from the literature to show where the results of this paper could be useful.

Abstract: Constraint satisfaction problems are ubiquitous in artificial intelligence and many algorithms have been developed for their solution. This paper provides a unified survey of some of these, in terms of three classes: (i) tree search, (ii) arc consistency (AC), and (iii) hybrid tree search/arc consistency algorithms. It is shown that several important algorithms, when slightly rearranged, are of the latter hybrid form, but with arc consistency components that do not necessarily achieve full arc consistency at the tree nodes. Accordingly, we define several new partial AC procedures, AC1/5, AC1/4, AC1/3, and AC½, analogous to the well-known full AC algorithms which Mackworth has called AC1, AC2, and AC3. The fractional suffixes on our AC algorithms are roughly proportional to the degree of partial arc consistency they achieve. Unlike traditional versions, our AC algorithms (full and partial) are presented in a parameterized form to allow them to be embedded efficiently at the nodes of a tree search process. Algorithm complexities are compared empirically, using the n-queens problem and a new version called confused n-queens. Gaschnig's Backmarking (a tree search algorithm) and Haralick's Forward Checking (a hybrid algorithm) are found to be the most efficient. For the hybrid algorithms, we find that it pays to do little arc consistency processing at the nodes, incurring more nodes, but sufficiently reducing the work per node so as to obtain less work over the whole tree. The unified view taken here suggests several new algorithms. Preliminary results show one of these to be the best algorithm so far. Les problemes de satisfaction des contraintes sont omnipresents dans le domaine de l'intelligence artificielle et bon nombre d'algorithmes ont eteelabores afin de les resoudre. Cet article fait etat d'une etude de ces algorithmes selon trois classes : les algorithmes de (i) recherche arborescente, de (ii) consistance d'arc (AC) et de (iii) recherche hybride arborescente – consistance d'arc. II est demontre que plusieurs algorithmes importants, lorsque legerement modifies, font partie de la derniere classe dite hybride, avec cependant des composantes de consistance d'arc qui ne permettent pas necessairement d'obtenir une consistance d'arc complete aux nœuds d'arbre. Par consequent, nous avons elabore plusieurs nouvelles procedures AC partielles, AC1/5, AC1/4, AC1/3 et AC ½, qui sont analogues aux algorithmes bien connus complets AC que Mackworth a appele AC1, AC2 et AC3. Les suffixes fractionnaux de nos algorithmes AC sont plus ou moins proportionnels au degre de consistance d'arc partielle qu'ils obtiennent. Contrairement aux versions traditionnelles, nos algorithmes AC (complets et partiels) sont presentes dans une forme parametree afin de leur permettre d'ětre emboites efficacement aux nœuds dans un processus de recherche arborescente. La complexite des algorithmes est comparee de facon empirique, a l'aide du probleme n-reines et d'une nouvelle version, dite n-reines confuses. Le marquage arriere de Gaschnig (un algorithme de recherche arborescente) et la verification avant de Haralick (un algorithme hybride) se sont reveles les plus efficaces. Dans le cas des algorithmes hybrides, nous avons constate qu'il etait profitable de peu traiter la consistance d'arc aux nœuds, car, bien que cela entraine davantage de nœuds, le travail par nœud s'en trouve reduit, au point de donner moins de travail pour l'arbre entier. Le point de vue unifie adopte ici suggere plusieurs nouveaux algorithmes. Les resultats preliminaires permettent de classer l'un d'entre eux comme le meilleur algorithme jusqu'a present.

Abstract: Many existing inductive learning systems have been developed under the assumption that the learning tasks are performed in a noise-free environment. To cope with most real-world problems, it is important that a learning system be equipped with the capability to handle uncertainty. In this paper, we first identify the various sources of uncertainty that may be encountered in a noisy problem domain. Next, we present a method for the efficient acquisition of classification rules from training instances which may contain inconsistent, incorrect, or missing information. This algorithm consists of three phases: (i) the detection of inherent patterns in a set of noisy training data; (ii) the construction of classification rules based on these patterns; and (iii) the use of these rules to predict the class membership of an object. The method has been implemented in a system known as APACS (automatic pattern analysis and classification system). This system has been tested using both real-life and simulated data, and its performance is found to be superior to many existing systems in terms of efficiency and classification accuracy. Being able to handle uncertainty in the learning process, the proposed algorithm can be employed for applications in real-world problem domains involving noisy data. De nombreux systemes ?apprentissage inductif existants ont eteelabores suivant le principe que les taches ?appren-tissage sont executees dans un environnement sans bruit. Afin de faire face a la plupart des problemes du monde reel, il est important que le systeme ?apprentissage soit en mesure de tenir compte de ľincertitude. Dans cet article, les auteurs identifient les diverses sources ?incertitude qui peuvent ětre rencontrees dans un domaine bruite. Par la suite, ils proposent une methode de saisie efficace de regies de classification pour des situations ?apprentissage pouvant comporter des informations incorrectes, incompletes ou incoherentes. Ľalgorithme comporte trois phases: (i) la detection de formes inherentes dans un ensemble de donnees ?apprentissage bruitees; (ii) ľelaboration de regies de classification basees sur ces formes; (iii) ľutilisation de ces regies pour predire ľappartenance ?un objet a une classe. La methode a ete mise en oeuvre dans un systeme appele APACS (automatic pattern analysis and classification system). Ce systeme a ete soumis a des essais en utilisant des donnees reelles et des donnees simulees; sa performance s'est averee superieure a celle de nombreux systemes existants en termes ?efficacite et de precision. Puisqu'il est en mesure de tenir compte de ľincertitude dans le processus ?apprentissage, ľalgorithme propose peut ětre applique a la resolution de problemes du monde reel qui mettent en cause des donnees bruitees. Mots cles: classification, donnees ?apprentissage bruitees, formes probabilistiques, regies probabilistiques, couplage partiel, poids de ľevidence, information mutuelle.

Abstract: A {\em truth maintenance system} is a subsystem that manages the utilization of assumptions in the reasoning process of a problem solver. Doyle''s original motivation for creating a truth maintenance system was to augment a reasoning system with a control strategy for activities concerning its non-monotonic state of beliefs. Hitherto, much effort has been invested in designing and implementing the concept of truth maintenance and little effort has been dedicated to the formalization that is essential to understanding it. This paper provides a complete formalization of the principle of truth maintenance. Motivated by Reiter and de Kleer''s preliminary report on the same subject, this paper extends their study and gives a formal account of the concept of truth maintenance under the general title of {\em assumption based reasoning}. The concept of assumption based theory is defined and the notions of explanation and direct consequence are presented as forms of plausible conclusion with respect to this theory. Additionally, the concept of extension and irrefutable sentences are discussed together with other variations of explanation and direct consequence. A set of algorithms for computing these conclusions for a given theory are presented using the notion of prime implicates. Finally, an extended example on Boolean circuit diagnosis is shown to examplify these ideas.