No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Explanation Based Generalization
... Si les contraintes sont munies de propagateur GAC, le principe ne dépend pas de l'ordre dans lequel ces propagations sont effectuées. Il est possible de générer avec un nombre linéaire d'étapes des ensemble en conflits minimaux au sens de l'inclusion[69] (ou encore logarithmique en procédant par dichotomie). Sachant que chaque ensemble en conflit engendrera au moins une violation, par une génération d'ensemble disjoints on peut obtenir une borne inférieure du nombre de contraintes violées. ...
La programmation par contraintes (PPC) a pour objet de résoudre des problèmes ayant une structure combinatoire, qui impliquent des contraintes de formes diverses. Les outils de PPC sont génériques et composables. Ils peuvent facilement être hybridés avec d’autres technologies. Depuis son apparition dans les années soixante-dix, la PPC a suscité un intérêt croissant dans la communauté scientifique. Une des raisons de son succès est qu’il s’agit d’un paradigme théoriquement déclaratif, comme l’avait souligné le Professeur Eugène Freuder : “Constraint Programming represents one of the closest approaches computer science has yet made to the Holy Grail of programming : The user states the problem, the computer solves it”. Ce pouvoir expressif permet de modéliser un grand nombre d’applications proches les unes des autres sans avoir à implémenter plusieurs fois leurs composants algorithmiques les plus critiques. La PPC est largement employée pour résoudre des problèmes industriels, au détriment parfois, paradoxalement, de sa déclarativité. Il peut en effet s’avérer nécessaire, voire indispensable, de concevoir la modélisation du problème en exploitant fortement des propriétés qui lui sont propres. Cet état de faits comporte le risque d’éliminer à tort certaines idées scientifiques importantes mais trop novatrices pour pouvoir être employées à court terme dans un cadre industriel. Le compromis entre généricité et efficacité se situe donc au cœur de la recherche scientifique en programmation par contraintes. L’objet de cette habilitation est de synthétiser un ensemble de tentatives de réponses à cette problématique, ainsi que les perspectives qui s’en dégagent. Je présente notamment une sélection de contributions portant sur les problèmes d’optimisation : traitement des problèmes sur contraints, de l’incertitude, ainsi que des outils conceptuels permettant une caractérisation des solutions plus fine qu’une notion d’optimalité fonctionnelle classique. Je mets en relief les motivations pratiques et théoriques de ces contributions, leur fil conducteur, ainsi que certaines questions qui demeurent ouvertes. En outre, j’évoque mes travaux sur l’aide à la modélisation de problèmes en programmation par contraintes, par apprentissage, et la génération de conditions de filtrage à partir de représentations génériques des contraintes globales.
... Another important area for future research is combining similarity-based and explanation-based methods. There has already been significant research on combining the two approaches along many different directions (Lebowitz, 1985; Mitchell, 1984; Pazzani, 1985; Porter & Kibler, 1985; Rajamoney et al., 1985). Finally, we wish to stress the important part that implementation of large computer systems played in our work. ...
In the last issue of this journal Mitchell, Keller, and Kedar-Cabelli presented a unifying framework for the explanation-based approach to machine learning. While it works well for a number of systems, the framework does not adequately capture certain aspects of the systems under development by the explanation-based learning group at Illinois. The primary inadequacies arise in the treatment of concept operationality, organization of knowledge into schemata, and learning from observation. This paper outlines six specific problems with the previously proposed framework and presents an alternative generalization method to perform explanation-based learning of new concepts.
... While the role of synchronous activity in the brain remains a matter of debate and controversy, a rich body of neurophysiological data suggests that such activity occurs in the brain and might even play a role in neural information processing (e.g., see Eckhorn, Bauer, Jordan, Brosch, Kruse, Munk, and Reitbock, 1988; Gray, Konig, Engel, and Singer, 1989; Murthy and Fetz, 1992; Abeles, Bergman, Margalit, and Vaadia, 1993; Singer, 1993; Singer and Gray, 1995; deCharms and Merzenich, 1996; Usher and Donnelly, 1998). Over the past few years, several models that also use synchrony to encode dynamic bindings during inference have been proposed (e.g., Park, Robertson, and Stenning, 1995; Sougne, 1996; Hayward and Diederich, 1997; Hummel and Holyoak, 1997). 5 Over the past ve years, shruti has been augmented in a number of ways in collaborative work between the author, his students, and other collaborators (see below). ...
We are capable of drawing a variety of inferences effortlessly, spontaneously, and with remarkable efficiency --- as though these inferences are a reflex response of our cognitive apparatus. This remarkable human ability poses a challenge for cognitive science and computational neuroscience: How can a network of slow neuron-like elements represent a large body of systematic knowledge and perform a wide range of inferences with such speed? The connectionist model shruti attempts to address this challenge by demonstrating how a neurally plausible network can encode a large body of semantic and episodic facts, systematic rules, and knowledge about entities and types, and yet perform a wide range of explanatory and predictive inferences within a few hundred milliseconds. Relational structures (frames, schemas) are represented in shruti by clusters of cells, and inference in shruti corresponds to a transient propagation of rhythmic activity over such cell-clusters wherein dynamic bindings a...
The inductive category formation framework, an influential set of theories of learning in psychology and artificial intelligence, is deeply flawed. In this framework a set of necessary and sufficient features is taken to define a category. Such definitions are not functionally justified, are not used by people, and are not inducible by a learning system. Inductive theories depend on having access to all and only relevant features, which is not only impossible but begs a key question in learning. The crucial roles of other cognitive processes (such as explanation and credit assignment) are ignored or oversimplified. Learning necessarily involves pragmatic considerations that can only be handled by complex cognitive processes.
We provide an alternative framework for learning according to which category definitions must be based on category function. The learning system invokes other cognitive processes to accomplish difficult tasks, makes inferences, analyses and decides among potential features, and specifies how and when categories are to be generated and modified. We also examine the methodological underpinnings of the two approaches and compare their motivations.
Game playing has attracted researchers in Artificial Intelligence ever since its beginnings. By comparison with human reasoning, learning by operationalization of general knowledge, as formalized by the Explanation-Based Generalization (EBG) paradigm, appears to be highly plausible in this domain. Nevertheless, none of the previously published approaches is (provably) sufficient for the target concept, and at the same time applicable to arbitrary game states.
We trace this paradox back to the lack of the expressive means of Negation as Failure in traditional EBG, and constructively support our claim by applying the respective extension proposed in [Schr96] to the chess endgame king-rook vs. king-knight.
Methodically, endgames are well-suited for quantitative evaluation and allow to obtain more rigorous results concerning the effects of learning than in other domains. This is due to the fact that the entire problem space is known (and can be generated) in advance.
We present the main results of a large-scale empirical study. The issues of training complexity, speedup for recognition and classification, as well as the question of optimal reasoning under time constaints are analyzed.
In this article we take a step towards providing an analysis of the Soar architecture as a basis for general intelligence. Included are discussions of the basic assumptions underlying the development of Soar, a description of Soar cast in terms of the theoretical idea of multiple levels of description, an example of Soar performing multi-column subtraction, and three analyses of Soar: its natural tasks, the sources of its power, and its scope and limits
In this paper, we study the generalization algorithms for second-order terms , which are treated as first-order terms with function variables, under an instantiation order denoted by � .F irst, we extend the least generalization algorithm lg for a pair of first-order terms un- der � , introduced by Plotkin and Reynolds, to the one for a pair of second-order terms. The extended algorithm lg, however, is insufficient to characterize the generalization for a pair of second-order terms, be- cause it computes neither the least nor the minimal generalization under � .S ince thetransformation rule for second-order matching algorithm consists of an imitation and a projection, in this paper, we introduce the imitation-free generalization algorithm ifg and the projection-free generalization algorithm pfg. Then, we show that ifg computes the min- imal generalization underof any pair of second-order terms, whereas pfg computes the generalization equivalent to lg under � . Nevertheless, neither lg, ifg nor pfg preserves the structural information. Hence, we also introduce the algorithm spg and show that it computes a structure- preserving generalization. Finally, we show that the algorithms lg, pfg and spg are associative, while the algorithm ifg is not.
Due to the difficult nature of Machine Learning, it has often been looked at in the context of "toy" domains or in more realistic domains with simplifying assumptions. We propose an integrated learning approach that combines Explanation-Based and Similarity-Based Learning methods to make learning in an inherently complex domain feasible. We discuss the use of explanations for Similarity-Based Learning and present an example from a program which applies thee ideas to the domain of terrorist events. L Introduction In this paper we present a novel approach to Machine Learning that integrates Explanation-Based and Similarity- Based methods to make learning in a complex domain feasible. We consider learning by observation in the domain of acts of international terrorism. The learning mechanism itself will be a component of a system that has the following task: Given information from a series of newspaper accounts of terrorist events, suggest an action that a law enforcement agency might take in response to those events. The domain of terrorist events is realistic and complex. Furthermore, since the input to the learning mechanism is limited to information from newspaper stories, we cannot assume that, in general, this information will be correct, complete, and consistent A typical description of a terrorist incident taken from the New York Times is: Paris, Feb 3. - An explosion, apparently caused by a bomb, ripped through a crowded shopping arcade on the Champs-Elysees tonight Eight people were wounded, three of them seriously... Witnesses said that damage was extensive.
In this paper a brief overview on the German Collaborative Research Center on Humanoid Robots is given. The research topics focused by the collaborative research center are multimodal man-machine interaction, learning of motion, skills and tasks, man-robot cooperation, perception and action and mechatronics and controls for a human like torus consisting of 2 redundant arms, 2 dexterous 5-finger hands, a head with eyes and ears, a neck and a spine. The project started in July 2001 and is planned in long terms for a period of up to 12 years. Over this period it is envisaged to elaborate and establish basic methodologies and technology for humanoid robots to act and cooperate close with humans in daily living environment. An integrated perspective towards continuous humanoid inter-action gaining multimodality, adaptivity, redundancy and flexibility is proposed.
Programming, cooperation and interaction with a humanoid robot is assumed to be multimodal which means to let the user program the robot simply by speech, gesture or demonstrating a task. The robot observes, interprets and then tries to imitate and to learn the performed user action. On the basis of a redundant humanoid type 2 arm robot system equipped with an active stereo head and a microphone array observation of the demonstrated task is realized as a first prototype using an active vision system. Grasping of objects is detected with the help of data gloves and active vision. The system interprets and stores the observed actions, segments them into meaningful sequences in a given context. Due to sensor errors and the complexity of the intended interaction, the system generates queries concerning intention, manipulation and objects. Motion catch techniques in combination with learning capabilities are combined towards imitation learning and active observational learning strategies.
Our learning system supports repeatable experiments in which practical problems related to document indexing and retrieval are addressed. Knowledge bases are typically critical to the document indexing and retrieval. In our experiments, one knowledge base is augmented with knowledge from another knowledge base, and at the first level this augmentation is an instance of similarity-based or analogy-based learning. But the overall system is tuned by interaction with a person so that changes in the representation of the knowledge base and in the performance and evaluation components of the learning system can lead to demonstrably improved performance. We provide evidence that small changes in structure or representation should correspond to small changes in function or meaning. Furthermore, all the components of the learning system depend in some way on detecting similarities and exploiting differences, and the move from similarity to analogy in learning exploits a second-order difference.
The paper provides an introductory survey of Explanation-Based Learning (EBL). It attempts to define EBL's position in AI by exploring its relationship to other AI techniques, including other sub-fields of machine learning. Further issues discussed include the form of learning exhibited by EBL and potential applications of the method.
Printout. Thesis (Ph. D.)--University of Illinois at Urbana-Champaign, 2003. Vita. Includes bibliographical references (leaves 104-113).
Implementations of Explanation-Based Generalization (EBG) within a logic-programming environment, as e.g. the well-known PROLOGEBG algorithm [KCMcC87], are able to generalize the proof of a goal from a definite (i.e. Horn clause) domain theory. However, it is a fact that practical applications frequently require the enhanced expressiveness of negations in rule bodies. Specifically, this is the case for the domain of game playing, where traditional EBG has turned out to be inadequate [Ta89]. In this paper we present an approach which extends EBG to this more general setting; it is described in the form of a transformation system, and comprises Siqueira and Puget's method of Explanation-Based Generalization of Failures [SiPu88] for definite programs. For the case that both domain theory and training example are represented as allowed normal programs, we prove that the derived clause satisfies the standard requirements of EBG, namely operationality, sufficiency, and correctness. Furthermo...
ResearchGate has not been able to resolve any references for this publication.