This paper presents the design and structure of the Finance Adviser system which is a knowledge-based consultation system. This system advises its users, usually owners of small-and medium-sized enterprises, on financial systems provided by governments for raising funds. A main function of this system is introducing suitable financial systems to users, which meet requirements and constraints set by users. This system includes the description of several types of knowledge with respect to financial systems and terms used in the process of consultation on fund raising, as its knowledgebase. That is, the knowledgebase consists of the descriptions of each financial system which will be an object of consultation and the descriptions of knowledge that are used for judgement whether a user meets conditions involved in some financial systems or not, for representation of funding aims of each system and for defining terms employed to organize a thesaurus. The thesaurus consists of two types of collections of terms. 0ne type is used by end users, many of them are usually ordinal civilians, the other is technical ones used by administrative specialists. Those two types of terms are co-related so as to achieve a better man-machine interface. That is, they are useful for seeking appropriate fiancial systems basing upon daily used terms which represent users' concrete desire for which the fund is spent, conventional terms used for naming of machine expected to be introduced into a factory, and so on. In essence, the thesaurus is constructed for narrowing semantical gaps between terms in daily language and terms in technical one by transforming the former into the latter. The technical terms are directly corelated with corresponding financial systems. By executing question-answering in daily language, the finance adviser system infers purposes of fund raising which are appropriate or relevant to find possible financial systems. Our experiments show that the system seeks suitable financial systems when measured using both recall-precision and usefulness.

Nowadays there are very many document databases. However, people often find it difficult to retrieve target documents from those databases. One reason for the difficulty is that keywords assigned to documents are not adequate. This paper presents a novel method for automatic keyword extraction from Japanese documents in a database. Conventionally, keywords have been extracted, based on various heuristics, with which the importance of individual words is measured. This paper proposes objective criteria for extracting keywords from a mass of candidate-words. They are efficiency criterion and recall criterion. The efficiency criterion concerns the efficiency involved in utilizing a word for retrieving a document from a database. The recall criterion for a word concerns the likelihood that that word is used as a keyword for database retrieval. Those two criteria are quantified statistically using distribution pattern of documents in a database. A product of the quantified criteria supplies a keyword-fitness measure for a word. Keyword extraction is implemented as an optimization of the keyword-fitness by Genetic Algorithm. An experimental result shows the validness of the keyword-fitness and suggets the complementarity of the authors' keyword-fitness and heuristics, when conventionally used.

There is qualitative reasoning as one method for explaining the causal relationship on the various physical phenomena with time factor. By using such qualitative reasoning, we can predict the whole behavior of a object, and make the system recognize its state. As a result of it, the system can generate the causal explanation related to its physical phenomena. In this study, we propose the architecture of the intelligent explanation system incorporated with the function of qualitative reasoning which treats pendulums colliding phenomena, as reflecting human naive thinking process. Especially, we consider to combine the qualitative reasoning system with numerous simulation which can make process mutually both the continuous state (connection of a moment and a boundary) and non-continuous state (connection of a moment and next moment) of the phenomena. Finally, we describe the experimental framework of an intelligent educational system with those functions and direct manipulation interface so called an interactive learning environment.

In this paper, we discuss a self organization mechanism of ALDB (Advanced Logical DataBase) that is capable of handling the complete-incomplete knowledge. A number of exceptional data will be included as a DB gets larger. The knowledge including exceptions can be considered as the incomplete knowledge, because it may derive inconsistency. For logical DB, the occurence of inconsistency will be crucial. Therefore, in order to maintain the consistency of DB, some self organization mechanism should be developed. We propose the self organization mechanism (SOM) of ALDB which is realized by transforming the complete knowledge into the incomplete knowledge. The ALDB consists of extensional DB (EDB), complete intensional DB (CIDB) and incomplete intensional DB (IIDB). EDB is a set of facts, whereas CIDB and IIDB are sets of rules representing complete knowledge and incomplete knowledge, respectively. The data retrieved from the ALDB is concerned with the ALDB extension that is a set of derived conclusions. The SOM is divided into the three steps : 1) the check of consistency, 2) the selection of a rule to be transformed and 3) the transformation to an incomplete rule and its ordering. When new data is added into the ALDB, the SOM firstly checks the consistency by investigating only the part of complete knowledge with the use of SLD resolution technique. If the ALDB is inconsistent, the SOM identifies a complete rule that is logically false by contradiction backtracking algorithm. The selected rule is then transformed to an incomplete rule whose intended meaning is "If…, then normally…". It allows the exceptions that are not satisfied with a complete rule. Finally, the SOM makes the ordering among the incomplete rules based on the generality of each rule. The ordering plays a significant role to reduce the number of ALDB extensions.

This paper proposes an inference algorithm for classification trees based on MDL, the Minimum Description Length Criterion. MDL is a well known model selection criterion which stipulates that one should select a model so as to minimize the total number of bits to encode information necessary to reproduce the original data given as input, including the description of the model and the parameters. First, an exact expression of the description length of classification trees with a certain encoding scheme is given. Then an efficient algorithm to generate classification treesis proposed, in which trees are produced in a top-down manner by reducing the coding length by replacements of a terminal node with a decision node. Finally, simulation experiments are described which show that our algorithm based on MDL outperforms that using another related model selection criterion called AIC. In particular, the classification trees obtained using MDL were much more consistent with the true model.

This paper introduces a new approach to Genetic Programming (GP), based on a system identification technique, which integrates a GP-based adaptive search of tree structures, and a local parameter tuning mechanism employing statistical search. In traditional GP, recombination can cause frequent disruption of building-blocks or mutation can cause abrupt changes in the semantics. To overcome these difficulties, we supplement traditional GP with a recovery mechanism of disrupted building-blocks. More precisely, we integrate the structural search of traditional GP with a local hill-climbing search, using a relabeling procedure. In an earlier paper, we introduced our adaptive program called "STROGANOFF" (i.e. STructured Representation On Genetic Algorithms for NOnlinear Function Fitting), which integrated a multiple regression analysis method and a GA-based search strategy. The effectiveness of STROGANOFF was demonstrated by solving several system identification (numerical) problems. This paper extends STROGANOFF to symbolic (non-numerical) reasoning, by introducing multiple types of nodes, using a modified "Minimum Description Length (MDL)" -based selection criterion, and a pruning of the resultant trees. The eflectiveness of this system-identification approach to GP is demonstrated by successful application to time-series prediction and to symbolic regression problems.

Generating and maintaining models of the design object is one of the desired features of intelligent CAD systems. To use a central model that represents the relationships among models of different aspects enables a modeling system to explicitly represent knowledge about physical phenomena. We call this central model in the model-based integration metamodel. In this study, a metamodel is represented in the framework of qualitative physics, among all, qualitative process theory. Reasoning about possible behaviors of the design object relates physical phenomena represented in different aspect models. The use of the metamodel is illustrated by an example of reasoning about an electromagnetic motor. Reasoning about relationships among quantitative models is also discussed through an example of finite element mesh generation.

In order to solve a complicated problem, we often simplify the problem into a more manageable abstract problem and obtain information for the original problem by solving it in an abstract form and by utilizing the abstract solution. Such technique is widely used in problem solving or planning. In the Conventional Planning system (ABSTRIPS), abstraction depends on the representation of rules (preconditions, deletion lists, addition lists) and only the precondition of the operator is abstracted by deleting atoms. In this paper we propose a new abstraction method based on homomorphism, where planning problems are formalized by logic programs on specialization system and are abstracted by homomorphism to abstract problems. Specialization system is the structure which consists of atoms and substitutions. Concrete problem or abstract problem is written by its own specialization system. We call a mapping from a concrete problem to a abstract problem 'homomorphism'. As compared with conventional planning systems, our method has advantages, especially regarding the following two points. The first is to be able to abstract not only the operators but also the states of the planning problems. The second is to have a homomorphism theorem, which provides the theoretica1 basis for the abstraction method in this paper.

Human beings accept external stimuli through the senses, and the information is transmitted to the brain. It seems that human beings acquire various concepts from these information. Particularly, auditory and vision stimuli play an important role in the concept acquisition. From this point of view, we developed a system which acquired concepts and grammar from visual and acoustic information without a priori knowledge by comparing input information with acquired concepts. And, the system generates sentences (or speech) for input images and image concepts for input sentence. A concept consists of a set of relation between a figure feature and a speech event. A figure consists of four categories i. e. shape, size, color and position, and speech events consist of speech waves which are extracted from input speech. If the system perceives the acquired concepts in the input speech, the learning process begins from mapping between the input information and the perceived concept. The part of input which was not spotted by acquired concepts is registered as the new concept in the concept dictionary. And, our system perceives the sequence of concept inputs by using acquired concepts, and models these sequences by using HMM, in other words, it acquires the grammar. Furthermore, our system generates an utterance which explains input images by using acquired concepts and grammar. Our system acquired 11out of 12 types of concepts from 100 pairs of utterance and images. Using the left-to-right HMM for grammar acquisition, the probability our system will generate correct sentences for input images is about 50%. We have realized the first stage of human's concept acquisition process on a computer system.

This paper discusses the important issue of knowledge base comprehensibility and describes a technique for comprehensibility improvement. comprehensibility is often measured by simplicity of concept description. Even in the simplest form, however, there will be a number of different DNF (Disjunctive Normal Form) descriptions possible to represent the same concept, and each of these will have a different degree of comprehensibility. In other words, simplification does not necessarily guarantee improved comprehensibility. In this paper, the authors introduce three new comprehensibility criteria, similarity, continuity, and conformity, for use with tabular knowledge bases. In addition, they propose an algorithm to convert a decision table with poor comprehensibility to one with high comprehensibility, while preserving logical equivalency. In experiments, the algorithm generated either the same or similar tables to those generated by humans.

We have proposed LMO (Local Minimum driven Organizing) as a group forming mechanism for agents that solve Distributed Constraint Satisfaction problems. It is summarized as follows. When an agent (A1) gets caught in a local minimum, 1) A1 sends its local CSP to the agent (A2) that shares a violated constraint, 2) A2 puts their CSP together and searches for all possible assignments with simple backtracking. One flaw of LMO is that the cost of search grows exponentially with the number of CSP being put together. To cope with this flaw we extend our method as follows and introduce the dynamically weight adjusting strategy : ・ We associate a weight with each constraint. The weights have 1 as their initial values ; ・ We measure the cost of instantiation as the sum of weights of violated constraints ; ・ Strategy : after putting CSP together, A2 sends neighbors the number of variables or the size of domain ; neighbors reassign it for the weights of constraints being shared with A2. The result of the experiment on distributed 3-coloring problem was that this strategy tended to protect against concentration of CSP and reduced the cost of search.

Students often fail to formulate phenomenon when they are solving dynamics problems. In the case of conventional ITS, their failures are tried to remedy in terms of dynamical concepts and formulas. However, the remediations are not always effective, because they often can't understand the errors including in their formulas. To correct the errors effectively, it is required to help them to recognize the errors. Generally, human is more sensitive to contravention of constraints in real world than in formula world. From this point of view, we propose a framework of "Fake Simulation" which can simulate the phenomenon reflecting to the errors including in the formulas. The students can recognize contravention of constraints in motion easier than that in formulas. Fake Simulation is one of visualization method of errors, which enables them to understand their errors and to correct the errors. Furthermore, it can prevent them recurring. We have already implemented Fake Simulation for dynamics. We also report the results of preliminary experiment for evaluating Fake Simulation.