前号 (Vol.9, No.4, pp.530-536) のAIマップは,我が国の自然言語処理分野のリーダシップを長年取ってこられた長尾 眞氏によるものでした.広い視野から現在の自然言語処理技術における問題点を論じる一方,これまでのアプローチに対しても厳しい批判を投じた provocative なものでした.ここではさまざまな観点から議論を沸騰させるために,自然言語研究においてそれぞれの立場の異なる4氏(大阪大学 郡司 隆男氏,電子技術総合研究所 橋田 浩一氏,東京工業大学 徳永 健伸氏,日本アイ・ビー・エム(株) 丸山 宏氏)にコメントをいただきました.また,コメントに対する長尾 眞氏の返答も同時に掲載することにいたしました. (編集委員会AIマップワーキンググループ)

Many AI tasks can be formulated as constraint satisfaction problems (CSP). Since CSP is NP-complete, it is expected that there are no effective algorithms for solving every CSP. So it is important to find out some simpler subclasses of CSP which can be solved in polynomial time. The paper proposes a new breadth-first algorithm for finding all solutions of a given CSP, and analyzes its complexity. According to the algorithm, a new hierarchy of subclasses of CSP computable in polynomial time, which we call k-bounded, is introduced. It is conjectured that determining the value k of a given CSP is NP-hard. By giving a concrete algorithm, however, it is proved that, when k is fixed to a constant, the computational complexity of deciding whether a CSP is k-bounded or not is O(n^<k+1>) in the best case, and O(n!) in the worst case. Some properties of k-bounded and k-tree are also clarified.

In this paper a systematic knowledge representation technique is described, by which different levels of a logical circuit can be smoothly represented and manipulated. We think that this knowledge representation model is a kind of deep knowledge and a cognitive modelling concerning an object system could be achieved. This modelling is done from two points of views that are the logical configuration model and the physical configuration model. The logical configuration model is represented associated to the logical behaviors of the object. For representing the behaviors we are combining the black box model and the gray box model. We call this way of representing a model an object model. In this paper relatively low functional levels of abstraction are dealt with, i.e., the logical level and the circuit level of a digital system. The physical configuration model represents an actual structure of the object. A four-bit full-adder circuit was used as an examples of the modelling. A trouble shooting system by means of the model-based reasoning is also described as an example of applications. The application system was developed using frame based knowldge representation language ZERO.

Many practical applications of diagnosis require the reliable identification of multiple faults of components and sensors in quantitative measures. However, the state of the art of diagnosis technique is considered to be still insufficient to meet these severe requirements especially for nonlinear and dynamic components and sensors. This research proposes a method to achieve these practical requirements by using the frameworks of optimized constraints, minimal conflicts based diagnosis, and causal ordering of physical systems. First, the detection of faulty behaviors of an objective system is performed based on the quantitative consistency checking between observations and the optimized constraints called as "minimal over-constraints". Second, once if some inconsistencies are detected, a mathematical operation of minimal conflicts based diagnosis derives the candidates of faulty mechanisms and functions even under multiple failure conditions. Third, the anomalous quantities directly disturbed by the faulty mechanisms are identified systematically based on causal ordering. Furthermore, the quantitative deviations of these quantities are evaluated by using the minimal over-constraints. The performance of the proposed method is demonstrated through an example to diagnose an electric water heater. The practicality of this diagnosis has been confirmed for the multiple failures in nonlinear and dynamic systems.

This paper presents a new type of multi-population GA, that is, the forking Genetic Algorithm (fGA). The fGA is designed to solve multi-modal problems which are difficult to solve by the traditional GAs because of the many isolated-optimums. The feature of fGA is that it uses multi-population consisting of one parent population with blocking mode and one or more child populations with shrinking mode. Two problems are taken to demonstrate the validity of fGA. One is an FM Sounds parameter identification problem and the other is Oliver's 30 City TSP. The results of experiments show that the fGA outperforms the standard GA.

In this paper, we propose a knowledge representation and reasoning scheme DOT^<*￢>, in which dot notations and regular expressions are used for representing IS-A relation as well as IS-NOT-A relation. Transferring a given knowledge expressed by regular expressions into an equivalent knowledge expressed without using regular expression, it is shown that the consistency of the knowledge can be decided in polynomial time. Since the system is closed under inference process in regular sets of dot notations, we can realize the reasoning mechanism defined on regular sets and incrementally specify a new knowledge in DOT^<*￢>.

Learning is defined as a process of obtaining results which contribute to improvement of the performance of problem solving. In most EBL systems, the efficiency of problem solving is adopted as a measure of the performance. This is the reason why utility problem is the most popular theme among EBL researches. From a practical point of view, however, it is necessary to consider not only utility but also usability of learning results. Although, in many cases, user's requirements are reflected on the usability measure, little effort has been devoted to usability research as compared with utility one. This is one of the reasons for EBL being behind in applicability to real world problems in comparison with SBL. Considering the fact that EBL is a framework which reorganizes the domain theory into useful one, it is desired EBL should cope with user's various performance measures which direct the reorganization process of the domain theory. In this paper we introduce some new measures of the performance, called comparison viewpoints, on which user's demand can be reflected into EBL. We formalize the refinement process of explanation structure including the generalization process of EBL, and improve the usability of learning descriptions by introducing comparison knowledge to that process. The applicability of EBL to real world problems will be accelerated by this extension.