Case-based reasoning is a recent approach to problem solving and learning that has got a lot of attention over the last few years. Originating in the US, the basic idea and underlying theories have spread to other continents, and we are now within a period of highly active research in case-based reasoning in Europe, as well. This paper gives an overview of the foundational issues related to case- based reasoning, describes some of the leading methodo- logical approaches within the field, and exemplifies the current state through pointers to some systems. Initially, a general framework is defined, to which the subsequent descriptions and discussions will refer. The framework is influenced by recent methodologies for knowledge level descriptions of intelligent systems. The methods for case retrieval, reuse, solution testing, and learning are summa-rized, and their actual realization is discussed in the light of a few example systems that represent different CBR approaches. We also discuss the role of case-based methods as one type of reasoning and learning method within an integrated system architecture.

Despite the fact that many symbolic and connectionist (neural net) learning algorithms are addressing the same problem of learning from classified examples, very little Is known regarding their comparative strengths and weaknesses. This paper presents the results of experiments comparing the ID3 symbolic learning algorithm with the perceptron and back-propagation connectionist learning algorithms on several large real-world data sets. The results show that ID3 and perceptron run significantly faster than does backpropagation, both during learning and during classification of novel examples. However, the probability of correctly classifying new examples is about the same for the three systems. On noisy data sets there is some indication that backpropagation classifies more accurately. 1.

Most work in artificial intelligence and law has concentrated on modelling the type of reasoning done by trial lawyers. In fact, most lawyers' work involves planning -- for example, wills and trusts, real estate deals, and business mergers and acquisitions. Certain planning issues, such as the use of underspecified, or "open-textured" rules, are illustrated especially clearly in this domain. In this thesis, I set forth the characteristic features of planning in law, place it in the context of past artificial intelligence work in both law and planning, and describe CHIRON, a system that I have developed implementing my theory of open-textured planning in the domain of personal income tax law.

1 This paper briefly reviews the two currently dominant paradigms in machine learning - the connectionist network (CN) models and symbol processing (SP) systems; argues for the centrality of knowledge representation frameworks in learning; examines a range of representations in increasing order of complexity and measures of similarity or distance that are appropriate for each of them; introduces the notion of a generalized distance measure (GDM) and presents a class of GDM-based inductive learning algorithms (GDML). GDML are motivated by the need for an integration of symbol processing (SP) and connectionist network (CN) approaches to machine learning. GDM offer a natural generalization of the notion of distance or measure of mismatch used in a variety of pattern recognition techniques (e.g., k-nearest neighbor classifiers, neural networks using radial basis functions, and so on) to a range of structured representations such strings, trees, pyramids, association nets, conceptual graphs...

Abstract. A key issue in case-based reasoning is how to maintain the domain knowledge in the face of a changing environment. During the case retrieval process in case-based reasoning, feature-value pairs are used to compute the ranking scores of the cases in a case base, and different feature-value pairs may have different importance measures, represented as weight values, in this computation. How to maintain a set of appropriate feature weights so that they can be used to solve future problems effectively and efficiently will be a key factor in determining the success of case-based reasoning applications. Our focus in this paper is on the dynamic maintenance of feature weights in a case base. We address a particular problem related to the feature-weight maintenance issue. In current practice, the feature weights are assigned and revised manually, not only making them highly informal and inaccurate, but also involving intensive labor. We would like to introduce a semi-automatic introspective learning method to partially address this issue. Our approach is to construct a network architecture on the case base that supports introspective learning. Weight learning and weight-evolution are accomplished in the background through the integration of a learning network into case-based reasoning, in which, while the reasoning part is still case based, the learning part is shouldered by a layered network. The computation in the network follows well-known neural network algorithms with well known properties. We demonstrate the effectiveness of our approach through experiments.

The simultaneous interpretation of object behaviour from real world image sequences is a highly desirable goal in machine vision. Although this is rather a sophisticated task, one method for reducing the complexity in stylized domains is to provide a context specific spatial model of that domain. Such a model of space is particularly useful when considering spatial event detection where the location of an object could indicate the behaviour of that object within the domain. To date, this approach has suffered the drawback of having to generate the spatial representation by hand for each new domain. An algorithm, complete with experimental results, is described for the automatic generation of a hierarchical region based context specific model of space for strongly stylized domains from the observation of objects moving within that domain over extended periods. The highest (hierarchical) level of region describes areas of behavioural significance or the paths followed by moving objects....

The paper presents an approach for improving the behavior of "traditional" knowledge-based systems (KBS) by integration with a special-designed case-based reasoner. An abstract classification KBS with a simple rule-based architecture has been chosen as a test-bed for illustration of how the expert knowledge may be corrected by experience accumulated during the real use of the system. The structure of the correcting module is based on the general architecture of a case-based planner adapted for solving classification tasks. The organization of the system case memory is based on a vocabulary of possible problem solving failures which may be caused either by the original KBS or by the case-based reasoner. 1. Introduction The recent experience in applying KBS in real domains have shown that the ability to solve difficult problems is tightly connected with the ability of the system to learn, since even the best domain theories are incomplete. So, the development of methods for improvement ...

ion and respecialization is a general structural adaptation technique that abstracts the piece of the retrieved solution, and respecializes it later. Respecialization results in applying other specializations of the abstractions to the current situation. Thus, it results in analogical problem solving. This technique is used in PLEXUS (Alterman, 1986) as a second step in adaptation process (first, a null adaptation is used, and only if something fails, the system replans it using abstraction and respecialization). A similar approach is used in PERSUADER (Sycara, 1987). 4. Critic-based adaptation is a structural adaptation based on using critics to debug almost correct solutions (Simmons, 1988; Hammond, 1989a; Gonzalez and Laureano-Ortiz, 1992). A critic checks if a particular combination of features can cause a problem in a plan. If such a feature is found, a specific repair strategy is applied for repair. In CHEF, several criticbased adaptation rules are used, e.g., deletion of unneces...

1 2 3 This paper describes a series of experiments with the cascade-correlation algorithm (CCA) and some of its variants on a number of real-world pattern classification tasks. Some of these experiments investigate the effect of different design parameters on the performance of CCA (in terms of number of training epochs and classification accuracy on the test data). Parameter settings that consistently yield good performance on different data sets are identified. The performance of CCA is compared with that of the backpropagation algorithm (BP) and the perceptron algorithm (PA). Preliminary results obtained from some variants of CCA and some directions for future work with CCA-like generative algorithms for neural networks are discussed. 4 1 Introduction Pattern classification and function approximation are two of the application areas for which several neural network architectures and learning algorithms have been proposed. The perceptron algorithm (PA) [Rosenblatt, 1958] is rathe...

Abstract not found