書誌事項
- タイトル別名
-
- Basic Design of Knowledge Compiler Based on Deep Knowledge
この論文をさがす
抄録
<p>Most expert systems built by the current generation tools use shallow knowledge (heuristics), which is a collection of "pattern & xrarr; action" associative rules without a deep understanding of the domain. Although the shallow knowledge approach contributed to the success of the first generation expert systems, it turns out to cause several problems. And it is suggested by several researchers that deep knowledge is the key idea to resolve these problems. Deep knowledge can be seen as the textbook knowledge which human experts have about the domain of the expertise. There is, however, little discussion on comparison of the cost for building shallow knowledge base with that for deep knowledge base. Namely, there is no idea about what kinds of deep knowledge an expert system should have and about how they are used in deep reasoning, which generates rules from various kinds of deep knowledge. Thus there is still some distance between the new generation tools based on deep knowledge and the current ones. This paper discusses Knowledge Compiler (KC), which is a system to perform deep reasoning in "machinery domain" aiming at developing a next generation tool. In regard with deep knowledge, we attach importance to the cost for acquiring it from the viewpoint of tools. For examples, since detailed causal network cannot be represented without domain-dependent form, it costs too much for coming into the tools and so we do not adopt such deep knowledge. Thus the authors have investigated deep knowledge as knowledge source for generating shallow knowledge, and obtained the following four kinds of knowledge: (1) Device World (DW) (2) Physical World (PW) (3) Control World (CW) (4) Interpretation World (IW) DW includes "intention of a machine designer" besides design information incorporating shape, measurements and configuration of components. PW has many kinds of physical principle and the condition on which it is applied. And it is useful in order to envision what phenomenon is caused in the machine. CW includes controllability of a physical parameter, durability of a component and observability of a symptom and then is used for rule generation process control. IW contains knowledge for mapping one physical state to a trouble-hypothesis and/or symptom. On the other hand, KC generates diagnosis rules from the above-mentioned four kinds of deep knowledge and then consists of Deep Forward Reasoning (DFR), Deep Backward Reasoning (DBR) and Control Modules. DFR tries to make propagation, starting from a given elementary symptom to get trouble-hypothesis relevant to it. DBR tries to make propagation, starting from the trouble-hypothesis obtained from DFR in order to get other symptoms and to accomplish the incomplete rule. CW is used to generate rules in order of "importance". Thus rule generation process is controlled by deep knowledge in CW. In this paper, diagnosing overheat problem of a car engine will be taken as an example. However, techniques developed here are common to all machineries and KC will be a kernel in developing the next generation tool.</p>
収録刊行物
-
- 人工知能
-
人工知能 2 (3), 333-340, 1987-09-01
一般社団法人 人工知能学会
- Tweet
詳細情報 詳細情報について
-
- CRID
- 1390285697603241856
-
- NII論文ID
- 110002807009
-
- NII書誌ID
- AN10067140
-
- ISSN
- 24358614
- 21882266
-
- 本文言語コード
- ja
-
- データソース種別
-
- JaLC
- CiNii Articles
-
- 抄録ライセンスフラグ
- 使用不可