It is difficult even for an expert to fully understand the behavior of a complex system. To do so, human beings often try to view such systems at different levels of abstraction based on their functional structure. The way we understand complex systems is thus mostly hierarchical. A method for hierarchical representation of complex system is proposed which enables acquisition and simultaneous utilization of knowledge expressed at multiple levels using different abstractions based on approximations. We can distinguish three stages on the path to automatic generation of a hierarchical knowledge base depending on the degree of automation. In this paper, the results of the first stage are described, including the implemented method as well as the hierarchical knowledge base constructed for digital circuits. The characteristics of this method can be summarized as follows : ・ Multiple domain theories and the relation between descriptions at different levels are used to express many aspects of a complex system. ・ To keep the whole hierarchical knowledge consistent, information about the approximation used in each hierarchy is stored. Without this information, contradictions which are introduced by the use of approximations decrease the functionality of the knowledge base. ・ An EBL-like method is used to support construction of the consistent hierarchical knowledge base complying with the proposed representation method. Also discussed is how the hierarchical knowledge base constructed by this method can be used in problem solving.

The crucial problem of hypothetical reasoning system is its slow inference speed, while it is a very useful framework in knowledge processing. We present a fast mechanism for the hypothetical reasoning, which uses the analogy of previous inference cases already proved to be true. An inference-path network can be effectively used for selecting useful hypotheses from analogical cases, and for generating new hypotheses which are necessary for proving a new goal. The inference speed of the hypothetical reasoning, whose computational complexity has been proved to be NP-complete or NP-hard, can not be improved from the exponential-order limit as long as we stay in ordinary search methods. This paper shows, however, that this limit can be improved to a large extent in average inference time by using analogy.

The basic idea of traditional similarity-based learning is that a program takes a number of instances, compares them in terms of similarities and differences, and describes the concept as a set of attributes common to positive instances. However, a concept exists because it is necessary to discriminate the concept from other concepts. Therefore, the all attributes common to positive instances are not always important to describe the concept. The important attributes are the ones which are necessary to discriminate the concept from other concepts. In this paper, I propose a new learning method based on differences among concepts. This method extracts the important attributes by changing the weight which is given to each attribute. Moreover, the method how the acquired concepts are memorized is important, especially when a given object is recognized or a concept is associated with other concepts. Therefore, a network structure based on similarity is proposed as a knowledge representation method.

Knowledge acquisition and verification is a critical bottleneck of expert systems. In particular, it is difficult to confirm and maintain the consistency of a large-scale knowledge base. Knowledge verification is the proceess that makes problem solving knowledge complete and consistenst. KNOV (Knowledge Verification System) has been developed for diagnostic applications. KNOV is a meta-system which regards the diagnostic knowledge as an assumption and verifies it by assumption-based reasoning using ATMS. An architecture with the following features is proposed for knowledge verification : (1) assumption-based reasoning with dynamic testing (2) meta-knowledge definition for verification (3) knowledge consistency using ATMS The validity and effectiveness of the ideas for knowledge verification were confirmed by applying KNOV to the diagnosis system of Electric Power Systems and Computer-Center Fault Recovery Systems used in the field. This paper shows a verification example using KNOV for the Electric Power System's diagnosis knowledge base. Last, problems to be solved for practical use of the system are pointed out.

This paper proposes a new learning mechanism called "production decomposition", which decomposes a single expensive production into several cheap ones. By the decomposition, instantiations of the expensive production are generated little by little. As a result, the number of instantiations that are actually calculated decreases in some cases improving the inference speed. Since the maximum number of instantiations during inference greatly decreases, the memory update cost of the state-saving-based pattern matchers such as Rete and Treat can also be reduced. There exists the optimal number of decomposites which minimizes the execution time. The decomposition, together with conventional composition, could optimize the size of individual production from the view point of inference speed. Related works, such as production composition, intermediate concept learning and the lazy matching algorithm are discussed to clarify the position of the proposed method.

The compilation of logical knowledge-base enables fast inference. The authors have reported a fast compilation algorithm named "Ordered Clause Consensus (OCC) method" which can compile propositional logic knowledge-base into prime implicant clauses efficiently. Since the expressive power is restricted in propositional logic, predicate logic with variables is required for the knowledge representation in many problems. This paper presents a compilation method for predicate logic knowledge-base as an extension of the OCC method. The OCC method for propositional clauses can be applied to the case of predicate clauses after converting the predicate clauses into their ground instances. Since this ground instantiation should be done in the Herbrand universe, its explosion becomes a serious problem. A compilation method namd "Ordered Predicate Clause Consensus (OPCC) method" of this paper uses the most general unification (MGU) to derive prime implicant clauses with variables. The technique of controling the sequence of resolution based on the number of biform literal pair is also incorporated to achieve the efficient derivation ofprime implicant clauses. In other words, this OPCC method directly applies the MGU and the OCC method to generate prime implicant clauses with variables, instead of applying lifting theorem to the prime implicant clauses generated by ground instantiation. The compiled knowledge-base consisting of these prime implicant clauses is particularly important for efficient abductive synthesis of hypothesis necessary to prove a given goal. In the case of generating prime implicant clauses, the OPCC method is 8～45 times faster than that of simple instantiation method including OCC method. When compared with breadth-first method for propositional logic knowledge-base, the OPCC method is 6～200 times faster for the predicate logic knowledge-base with the same content. Experimental study shows that the effectiveness of the OPCC method over simple instantiation method becomes larger as the scale of knowledge-base increases.

In this paper, we try to describe dynamic transition process in novices' computer programming based on goal/plan analysis. First, we design psychological experiments, in which subjects' programming process is recorded by both a tape recorder and a video camera. Secondly, we analyze subjects' semantic protocols, and convert the programming process into a series of scenes, each of which corresponds to a subject's thinking state. Thirdly, each of scenes is labeled based on categories established in advance. Through analysis of a series of labeled scenes, we identified six kinds of transition patterns in programming process, that is, (1) S-Process where syntactic error is recovered, (2) M-Process, semantic error is recovered, (3) T-Process, knowledge is acquired through tutoring, (4) F-Process, some different procedures appear one after the other without interaction with external environment, (5) R-Process, correct procedures which were applied previously are reinspected and (6) N-Process, some irrelevant procedures are applied continuously. We can explain successfully the regularity in novices' computer programming based on the six kinds of process. Moreover, through detailed observations of the six kinds of process above, we explicate the state of novices' knowledge on computer programming and acquisition and elaboration process of the knowledge. That is, when there are no relevant plans to solve a certain goal in subjects' knowledge base, it is pretty difficult for novices to construct a new plan by their own effort. In this case, tutor's instruction is an important factor to acquire the knowledge. Even when there are relevant plans, the plans have a lot of errors and large ambiguities, so novices must correct the errors and reduce the ambiguities. In this case, some kinds of messages from computer system : for example, compile error message, erroneous results of computation, become more prominent than tutor's instruction to correct the knowledge. We explain the role of those kinds of interactions with external environment (tutor's instruction, messages from computer system). We also explicate developmental process with continuous learning, where novices acquire more plans to solve a lot of goals, reduce ambiguity of the plans, and establish the correct knowledge base related to computer programming.

This paper describes design philosophy of a knowledge acquisition system named KAISER (a Knowledge Acquisition Inductive System driven by Explanatory Reasoning). This system is an intelligent workbench for constructing knowledge bases for classification tasks. As the acronym implies, KAISER first learns classification knowledge inductively from examples given by human experts, then analyzes the result based on abstract domain knowledge which is also given by the experts, in order to detect some unsatisfactory and/or unreliable conditions called improprieties. By interpreting the improprieties, this system invokes intelligent interview for acquiring new knowledge to eliminate the improprieties. The interview stimulates the human experts and help them to revise the learned results, control the learning process and remind new examples and domain knowledge. Viewed in AI aspect, this process provides reasonable motivation for interview by integrating similarity-based inductive learning and explanation based deductive reasoning based on the concept of improprieties. This approach makes it possible to compensate for the problems of inductive learning, explanation and knowledge acquisition interview each other. A simple example of diagnosis problem is also provided to clarify and evaluate the knowledge acquisition process.