前号(Vol.9, No.2, pp.233-240)のAIマップは,いろいろと話題が豊富で,またAI研究分野にも多大な影響を与えた第五世代コンピュータプロジェクトの中心人物,渕 一博氏によるものでした.このプロジェクトについてはすでに数多く語られていますが,プロジェクトは氏の考え方,個性を強く反映したものであったことから,氏の個人的な気持ちを素直に語っていただいたことで,より内面的な理解が得られたのではないかと思います.プロジェクトの評価については国内外でさまざまな意見があり,当然氏の意見についても異論・反論があると思います.ここでは3氏のコメントを掲載させていただきます.(編集委員会AIマップワーキンググループ)

There are many models and laws in sciences, which are qualitative in the sense that emphases are laid on rough-cut relationships among attributes of the objects of the studies rather than detailed qualitative models. Qualitative models are prevailing in social sciences such as economics and management sciences, explanation of behavior of devices, educational domains, thought experiments, and so forth. In many cases, relations among attributes are derived from those among attribute values. Functional dependency and order dependency in relational database theory are a typical example of that kind of relations. The main objective of this paper is to investigate basic logical validity, namely, soundness and completeness, of reasoning based on some qualitative models of order-dependency type : Two types of order dependency and two variants, namely, OD-I,II and TOD-I,II, are defined. Two ODs are derived from families of partial orders between attribute values, and TODs are derived from those of total orders. Inference axioms are provided for ODs, while definitions of TODs are translated, in a certain manner, into propositional logic, and thereby logical validity is proved.

The utility problem has appeared through the intensive research on EBL (Explanation-Based Learning) in recent years. The problem means a trade-off between the benefit of learned knowledge and the cost to locate and apply the knowledge. One of the basic solutions is a search for a benefit and harmless subset of learned knowledge, however most of the searches are very expensive and thus need the off-line evaluation to ignore the learning cost. This paper describes an inexpensive method to search a locally optimal macro-ordering in EBL, macro-operator learning. The complexity of the search is restricted within linear order of the amount of learned macros, and the necessary information including statistics on given problems is obtained through the execution of a simple EBL procedure through partial problems. For verifying the effectiveness of our approach, we implemented an IS-EBL using the inexpensive estimation of the utility and made on-line evaluation with a non-learning system and a non-selective EBL in a map-coloring problem. As a result, we had good experimental results and clarified a class in which the IS-EBL works well.

We have developed a health evaluation system on the base of knowledge engineering. The system, called HELSMEK (Health Care Support System with Medical Knowledge), consists of a personal computer, an OCR (Optical Character Reader) and a printer. Questionnaire data about the past history, family history, present symptoms and etc. are entered by OCR sheet. Findings of physical examination, X-ray, electrocardiograms and etc. are entered by OCR. Laboratory data of blood and urine are inputted by FD. The system automatically selects possible diagnosis and proper advices by using medical knowledge base. The knowledge base of the HELSMEK have been developed by four medical doctors, whose specialites are different, using the easy-to-use knowledge base editor. The system has been put into practical use in 4 hospitals and 15 health care centers of NTT to support the works of health care personnel. Through the practical operation of the system in one hospital, we have had the following results. (l) Up to 93.0% of the diagnosis produced by the system were coincident with those made by medical doctors. (2) The 95.7% of the advice sentences selected by the system were judged by the doctors to be unnecessary to correct. (3) To 198O examinees, 1632 different combinations of advice sentences has been given by the system.

To cope with the knowledge acquisition bottleneck, the authors propose a new architecture combining RBR (Rule Based Reasoning), CBR (Case Based Reasoning) and knowledge acquisition technology in a system which solves pattern search problems. The RBR part searches for specified patterns in a large space represented by a network structure such as an LSI circuit diagram, which contains a great number of patterns and variations. It then carries out specified actions, such as fault diagnosis, on the patterns that are found. The outputs of the RBR part are transferred to the CBR part. The user of the system detects and repairs a few pattern detection errors caused by the RBR part. The CBR part detects and repairs all remaining errors which can be estimated from the user detected ones. The repaired results are sent back to the RBR part to recover the RBR output. The repaired results are also stored automatically in the case base. Similar cases are grouped in a same case family. The knowledge acquisition part relates each case family to an incomplete rule in the RBR knowledge base and proposes modifying the rule. Eventually, the system can get refined rules with the cooperation of domain experts. Thus, the problem solving process and knowledge acquisition process are performed cyclically. The architecture was successfully applied to a pair condition extraction problem for an analog LSI circuit layout system.

Cross-modal metaphors are based on some "view-relations" between concepts in different domains, i.e. knowledge about viewing a mental (or abstract) concept as another Physical concept. Our goal is to establish an effective method of computationally eliciting those pieces of knowledge from cross-modal metaphors we human beings use in everyday life. Since we understand that analogical mappings between two non-identical concepts underlying cross-modal metaphors represent viewing-relations, we proposed a method for determining analogical mappings of metaphors as follows. Basically, we employ the structural consistency constraint in determining analogical mappings which allows the mappings of non-identical two relational predicates. In general, however, only this simple version of structural consistency is too weak as a constraint and tends to produce so many meaningless alternatives of analogical mappings. This paper addresses the problem by employing a semantic constraint, celled primitive matching, only at the initial stage of finding analogical mapping ; based on the representation that verbal concepts are described by use of abstract primitives only, the program first searches for the pairs of first-order literals with exactly the same Primitive Predicates which yield some initial mappings. And after that, the program searches for the whole correspondence between two concepts which does not contradict the initial mappings, based on the simple version of structure consistency constraint. In this paper, we made an experiment of analizing some cross-modal verbal metaphors by the primitive matching approach, as a result of which we show that the approach proposed is more advantageous in cpu-time saving than the simple version of structural consistency approach in some classes of verbal metaphors, and we also illustrate the obtained knowledge about viewing mental concepts as physical ones, which is general enough for use in interpreting other metaphoric expressions.

The purpose of this study is to make CORES (Contextual and Relative Existence System), which can learn conceptual relationships based on only a small number of instances under various contexts and relativity, as the first step of making an artificial intelligence model of human natural language processing by applying it first to pronoun resolution and inference of ellipses. Based on many psychological findings concerning pronoun resolution and ellipsis inference, it was hypothesized that early incremental interpretation should be necessary, that evaluation should be based on lexical information, temporal proximity, and verbs' relational information, and that continuation of every process to the end should be indispensable. These hypotheses were first tested and supported by a psychological experiment with 84 subjects. CORES was then improved based on the experimental results and adopted some new components and variables such as affinity nodes, temporal proximity nodes, psychological distances, and activation levels. Finally, the validity of new CORES was confirmed by simulations using the experimental materials. These results were discussed in terms of the necessity of utilization of world knowledge or common sense and that of further exploration and introduction of heuristics.

The concept of inheritance is one of the most important notions in knowledge representation paradigm and the object-oriented programming/databases. However, we have several issues of computation for handling the inheritance mechanism which include the infiniteness property of the domain as well as the procedure of computation. To overcome this deficiency, we propose a framework for knowledge representation called DOT, which employs the regular expression for representing IS-A relations defined on the domain extended by dot notations. In this DOT, introducing a transformation between two equivalent knowledge representations, the possibly infinite set of all answers to a query can be finitely specified using regular expressions. In such a manner, the problem of infiniteness is easily and directly handled under the framework of DOT, and this DOT is closed in the class of regular sets regarding the computation of inference based on its inheritance mechanism.