In this paper, we discuss the role of clustering techniques in the design of neural networks. Specifically, we address the issue in relation to two network paradigms: one based on back-propagation and the other based on radial basis functions. In the former case, we demonstrate, emprically, that by employing clustering techniques, the training effort may be drastically brought down. In the latter case, we demonstrate that clustering techniques can be employed to build more robust classifiers. We also discuss the role of clustering in the design of hierarchical systems. Specifically, we discuss a hierarchical system based on radial basis functions.

Clustering is concerned with grouping a collection of input objects. Conventional clustering algorithms cluster unlabelled objects. We argue that there are useful applications that involve clustering of labelled objects. We propose an approach for clustering of labelled objects. The proposed approach makes use of the domain knowledge represented in the form of a directed acyclic graph for clustering. We also propose a set of proper axioms in logic as a basis for the proposed algorithm. We study some of the properties of the approach such as order-independence and describe in detail an application of the proposed algorithm in the context of document retrieval.

Clustering techniques have been used for data abstraction. Dara abstraction has many applications in the contect of data-bases. Conceptual models are used to bridge the gap between the user's view of a database and the physical view of the database. Semantic models evolved to overcome the limitations of classical data models such as network and relational models. The paper uses a knowledge-based clustering algorithm to extend the abstractions, such as classification and association, which are employed in the semantic modeling of databases. The complexity of the proposed clustering algorithm is analysed. The extended semantic model can be used to design databases in which useful and interesting queries can be answered. The efficacy of the proposed knowledge-based clustering approach is examined in the context of a library database.

Belief revision systems aim at keeping a database consistent. They mostly concentrate on how to record and maintain dependencies. We propose an axiomatic system, called MFOT, as a solution to the problem of belief revision. MFOT has a set of proper axioms which selects a set of most plausible and consistent input beliefs. The proposed nonmonotonic inference rule further maintains consistency while generating the consequences of input beliefs. It also permits multiple property inheritance with exceptions. We have also examined some important properties of the proposed axiomatic system. We also propose a belief revision model that is object-centered. The relevance of such a model in maintaining the beliefs of a physician is examined.

We describe a clustering technique that can exploit a large body of knowledge. An algorithm to cluster input objects, using the knowledge available, is presented. This algorithm is order-independent and can be naturally extended to cluster objects in an incremental way. The usefulness of the knowledge-based clustering algorithm is studied in the context of a database.

Database comparison is an important area of research. Comparison of databases is at two levels: data level and structure level. In this paper, we investigate the applicability of the conventional clustering algorithms for the purposes of database comparison at both the data level and the structure level. We propose a knowledge-based clustering approach for structure level comparison. The above algorithms have been implemented in LISP and are used for the comparison of library databases.

The paper deals with a model-theoretic approach to clustering. The approach can be used to generate cluster description based on knowledge alone. Such a process of generating descriptions would be extremely useful in clustering partially specified objects. A natural byproduct of the proposed approach is that missing values of attributes of an object can be estimated with ease in a meaningful fashion. An important feature of the approach is that noisy objects can be detected effectively, leading to the formation of natural groups. The proposed algorithm is applied to a library database consisting of a collection of books.

This paper deals with a “default” learning scheme. The closely related paradigm is learning by observation. As inductive inference forms the basis for the proposed learning scheme, the default generation process acquires nonmonotonic behaviour. This nonmonotonic behaviour is captured by using a formal system with a modified version of the modus ponens as the inference rule. We propose an intuitively-appealing classification of defaults into four categories. The classification is essentially based on specific and general defaults. The formal system, presented in this paper, deals with specific defaults that form the basis for general defaults. We also observe that the bidirectional defaults can be successfully employed in reasoning.

The general form of a decision-making model raises hypotheses about the dynamics of stimulus, conception, and response. One might suppose that the decision progress begins with the perception of some sort of stimulus. The amount of information available for decision-making is often incomplete and hence any logic that attempts to model decision-making must be nonmonotonic in nature. The existing logics can handle defeasible nonmonotonic inferences. The authors propose a modified first-order logic so that defeasible beliefs can also be handled. The modification is in the form of a set of proper axioms to handle belief revision, and a modified modus ponens to capture nonmonotonic reasoning. The proposed logic is extended to model decision-making activity.<>