To represent human knowledge in a formal system, we have to deal with vagueness, which is further divided into vagueness of the knowledge holder and vagueness of the knowledge itself. We call the former confidence factor and the latter subjective probability of the knowledge. Traditionally, these are represented as a single parameter, called an uncertainity factor, in most of expert systems. We introduce a Probability Belief Operator (PBO) which is the combination of two factors. PBO represents higher order probabilistic belief. By providing axioms on PBO, we show that the system is a generalization of certainity factors such as used in MYCIN. We will provide inference rules over PBO so as to use the system for inference dealing with belief and uncertainity. We show some examples of probablistic inductive inferernce. We then show that PBO is a generalization of belief axiom. The formalism is also applicable to knowledge acquisition.

This paper is concerned with the construction of a knowledge system for "Risk Assessment and Risk COMmunication (RARCOM)". Such a system deals with health risks associated with drinking water contaminated by various hazardous materials in the Tokyo metropolitan area. In order to evaluate where and how much health risks associated with drinking water are distributed, we use the "Object Model" that describes the structure and the behavior of an urban water network. This object model is represented as a "component-conduit model", in which the structure of the object is represented as a "device topology" of a network of components and conduits, the behavior of which is defined as a set of behaviors, wherein the hydrological, bio-chemical and technical reactions and treatments of each components are included. Since the risk problem in an urban water network is related to multiple domain knowledge, it is insufficient to describe the object model of the urban water network using only a single ontological scheme. Thus this system employs a multiple ontological scheme for mapping the real system of urban water network onto the component-conduit model of the object. Using the object model of urban water network, RARCOM can assess the magnitude and the distribution of risks under various situations that are presented by various "what-if" scenarios.

This paper propose a theoretical framework for a knowledge programming system based on Reiter's default logic, where knowledge including complicated exceptions can be expressed simply as a default without enumerating exceptions explicitly. For this framework, we introduce a default theory called a normal Horn default theory and two kinds of constraints for the extensions of the default theory, called negative and positive constraints. Knowledge is represented using the normal Horn default theory and the two constraints. An inference of a default theory is usually very inefficient because of its non-monotonicity. To resolve this situation, we also give a program transformation method that incorporates the constraints into the normal Horn default theory. An inference of a normal default theory is achieved monotonically. Thus the program transformation yields an optimized knowledge program.

The process of concept learning can be viewed as a search for general concept descriptions that cover all of the positive examples and none of the negative examples given by the teacher. In such learning, the examples give crucial influence on the efficiency of learning. However,it is not easy for the teacher to select good examples. On the other hand, the learner can select effective examples for verifying the candidates of the result, because he will always know the descriptions as the candidates. Accordingly, the way that the learner selects examples and questions whether each of them is positive or negative is of great importance. Such learning is called, in general, interactive concept learning. In interactive concept learning, there are two factors to realize the efficient learning. One is the number of questions, and the other is the cost for selecting an example. So far, there have been proposed various method of interactive concept learning. Halving guarantees to learn by the minimum number of questions, and it requires high cost for selecting an example. Conservative Selection (CS) requires low cost for selecting an example, and the number of questions increases in case that a lot of generalization is needed. Middle Hypothesis Selection (MHS), which we have proposed, requires low cost like CS, and the number of questions increases in case that the generalization tree has a large number of branches. In this paper, we propose a new interactive concept learning method, an improved version of MHS, named Adaptive Hypothesis Selection (AHS). AHS dynamically applies two ways for selecting examples by the criterion function, which determines relevant hypothesis selection based on the number of middle hypothesis and the depth of generalization tree. We discuss the efficiency of AHS compared with the existing methods and show that it can learn in less number of questions. And we experimentally verify the validity of the proposed criterion function.

Recently developed production system techniques have improved the performance of problem solving several times over. However, these techniques cannot be solutions for real-time continuous problem solving systems, where meaningful results are required within a predefined time limit. To satisfy real-time constraints, this paper proposes an organizational approach, where problems are solved not by a single production system, but by an organization of distributed production system agents. To explore the effectiveness of the organizational approach, we are studying the adaptive load balancing problem in which multiple problem solving organizations share a collection of processor resources. Problem solving requests arrive at the organization continuously, at variable rates. Meaningful results are required within a (possibly changing) time limit. Since the organization shares computing resources with other organizations, it does not suffice to decompose to maximal parallesism -the collective must adapt itself to take advantage of resources it can, but it must also adaptively free up resources for others while continuing to operate. Organizational self-design is then performed to equalize response times. For this purpose, we introduce new reorganization primitives, composition and decomposition, which change the number of agents in organization. Various simulation results prove the effectiveness of composition/decomposition techniques for building adaptive real-time production systems.

Explanation-Based Learning (EBL) fails to accelerate problem solving in some problem domains. An EBL system therefore needs to evaluate the utility of EBL to a given problem domain for determining whether it adopts EBL. Conventional EBL systems empirically evaluate the utility through test experiments solving a great number of test examples, which results in high evaluation cost. These empirical methods prohibit us from estimating the utility before conducting the test experiment. This paper presents a formal framework in which the utility of EBL, implemented in a logic programming, is computed analytically without any test experiments. We represent the utility of EBL as a function of two variables, the number and the distribution of test examples, and predict the utility on subsequent problems by analyzing the function. The utility function is determined by analyzing the trace of problem solving on training examples, not test examples. First, we define a standard EBL procedure including both problem solving and how to add learned rules to a rule base. After only training examples were solved, our method can compute the utility of EBL by evaluating computational cost both of an EBL system and a non-learning system assuming a distribution of test examples. Since our method needs no experimental result on test examples and no execution of learning procedures like EBG, it can efficiently predict the utility of EBL before test experiment. The utility of EBL obtained by our method is a function of the number of test examples, and it approaches to the utility evaluated empirically thus far as the number increases. Finally we show examples of computing the utility of EBL in a Mitchell's SAFE-STACK example. As a result, we have very interesting results that EBL deteriorates problem solving even in simple domain theory such as safe_to_stack (X, Y).

This paper describes a way of organizing large knowledge base which aims at human use in book form and in hyper-text system. This is based on the experience of the author when he was the chief editor of the Encyclopedia of Computer Science published by Iwanami Book Co. The book includes about 4500 entry terms and about 13000 index terms. The total pages are about 1200, in which about 800 pages are for the description of entry terms, 100 pages for a hierarchical tree representation of entry terms, and 300 pages for KWIC index table. The paper describes how the terminological words were systematically collected, and how these terms were organized into a hierarchical tree structure. Then the way of giving definition for each term is described in detail. Intensional definition, extensional definition, functional and other definitions are to be given by considering always the similarity and distinctness to other related terms. Finally the access methods to the information are described. Users have not necessarily an exact term to consult the encyclopedia. KWIC-form organization of index terms based on radix of compound terms is quite useful for such user situations. The hierarchical tree organization of terms is useful for the understanding of the whole structure of the area of computer science, and for the access to what users want to know. There are several other access methods provided for in the description part of each term.

Learning of large-scale neural networks suffers from computational burden and the local minima problem. One way of avoiding these difficulties is the adoption of module structured networks. Proposed in this paper is a learning of module structured networks using the structural learning algorithm with forgetting of link weights previously proposed by the author. The essence of the previous proposal is that the forgetting of link weights makes unnecessary links fade out, thus generating a skeletal network structure. The structural learning enables formation of modules. It also enables automatic utilization of the most appropriate module among the previously learned ones for the learning of the present module. This not only achieves efficient learning, but also makes the resulting network understandable due to its modular character. In the learning of a Boolean function, a module corresponding to its subproblem is automatically selected from among the previously learned modules. When there exist more than one modules corresponding to the subproblem, the best learned one is selected. In the parity problem, the structural learning can automatically use the result of the previously learned parity problem modules of the lower order. The transformation of figures can be represented as a sequence of basic transformations. Suppose each module has already learned one of the basic transformations. The structural learning can discover a sequence of modules which realizes the given transformation. These examples demonstrate the effectiveness of the structural learning with forgetting in the learning of module structured networks.

This paper describes a method of automatic hypertext construction from an Encyclopedic Dictionary of Computer Science, which was originally published in a book form. By checking varieties of sentential styles, sentences which define the meaning of head words and other important words are first extracted from the dictionary. These include the sentences expressing synonyms and giving intensional and extensional definitions. Then these sentences are further analyzed and the keywords and their relations are extracted automatically by relatively simple natural language processing. These extracted terms which include head words and index terms are connected each other according to the extracted relations from the defining sentences. The whole text data which amounted 2 mega characters were stored in a workstation. We constructed a user-friendly interface on X-window system, which realizes varieties of accesses to required information. The system includes full text search facility which scans 4 mega bytes data in a few seconds. This method of automatic hypertext construction is applicable to other encyclopedic dictionaries.