Naomichi Sueda

Knowledge acquisition in image processing expert system 'EXPLAIN'

A knowledge-acquisition facility has been incorporated into an expert system called EXPLAIN, which assists the nonexpert in using a package of image-processing algorithms to obtain a required image from a given one. It captures knowledge of image-processing procedures through interaction between the domain expert and the system. The authors discuss the method and its features: stimulative interview and rule generalization. A stimulative way of interviewing using analogical reasoning reduces the difficulty for the expert to elicit deeply and unconsciously embedded knowledge. The process of decomposing a successful sequence of image processing by the expert creates generalized rules in a natural manner.<<ETX>>

Use of multiple cases in case-based design

A framework of case-based design and its application to the mechanical design of induction motors are described. In the proposed framework, the use of multiple cases is essential in reducing the problem-solving complexity and ameliorating the knowledge acquisition bottleneck. Two subproblem generation processes are introduced: case-based reduction and goal-directed decomposition. In order to implement these processes effectively, a design model is introduced which maps requirement differences to the part of the design to be modified, and a decomposition algorithm is applied which takes into account the retrieval possibility and interaction occurrence. In a preliminary evaluation, the motor design systems performance was found to be as good as that of a human designer.<<ETX>>

A pilot system for plant control using model-based reasoning

Conventional heuristic-based expert systems for diagnosis and control cannot adequately handle unforeseen abnormal events. Our system combines heuristics with model-based reasoning to control a thermal power plant, and to diagnose and resolve abnormal events at the plant. >

Model-based diagnosis using qualitative causal model and set-covering

A model-based diagnostic system is proposed for continuous physical devices such as a thermal power plant. The aim of model-based diagnosis is to find faulty components in the model from observations. Set-covering is an approach to diagnosis when the causal relations between symptoms and disorders are clearly defined. The new method combines the model-based approach and set-covering approach. The authors introduce the qualitative causal model (QCM) and define symptoms and qualitative disorders in QCM. The system has two submodules to calculate diagnosis from symptoms. The qualitative propagation module calculates the causal relations between the symptoms and the qualitative disorders. The BIPARTITE module based on set-covering generates all the diagnoses using these relations. The proposed method is proved to realize the diagnosis from first principles in the continuous physical domain.<<ETX>>

Information Space Optimization with Real-Coded Genetic Algorithm for Inductive Learning

New feature construction methods are presented. The methods are based on the idea that a smooth feature space facilitates inductive learning thus it is desirable for data mining The methods, Category-guided Adaptive Modeling (CAM) and Smoothness-driven Adaptive Modeling (SAM), are originally developed to model human perception of still images, where an image is perceived in a space of index colors. CAM is tested for a classification problem and SAM is tested for a Kansei scale value (the amount of the impression) prediction problem. Both algorithms have been proved to be useful as preprocess steps for inductive learning through the experiments. We also evaluate SAM using datasets from the UCI repository and the result has been promising.

Plant control expert system coping with unforeseen events—model based reasoning using fuzzy qualitative reasoning

An ordinary expert system controls a plant according to heuristics. So, it fails to control the plant for lack of heuristics if unforeseen events occur as a result of abnormal situations. We propose a new framework of model-based reasoning that can dynamically generate the knowledge for plant control against unforeseen events. This proposed framework consists of three functions: (a) generation of the goal state after recovery from the unforeseen events; (b) generation of knowledge for plant control; (c) prediction of process trend curves and estimation of the generated knowledge. In the proposed framework, various kinds of models which correspond to the fundamental knowledge about plant control are used. We have implemented a thermal power plant control expert system on the basis of this proposed framework. This paper describes the model-based reasoning mechanism of the experimental plant control expert system to realize each of three functions. Especially as for (c), this paper explains qualitative reasoning mechanism using fuzzy logic.

Knowledge engineering application in image processing

To establish the logic and procedure for image processing, a specific software module must be selected from 350 software modules, named SPIDER in Japan, according to requirments and the original image. Highly specialized and experimental knowledge is required to select this software module.

Information Space Optimization for Inductive Learning

New feature construction methods are presented. The methods are based on the idea that a smooth feature space facilitates inductive learning thus it is desirable for data mining. The methods, Category-guided Adaptive Modeling (CAM) and Smoothness-driven Adaptive Modeling (SAM), are originally developed to model human perception of still images, where an image is perceived in a space of index colors. CAM is tested for a classification problem and SAM is tested for a Kansei scale value (the amount of the impression) prediction problem. Both algorithms have been proved to be useful as preprocess steps for inductive learning through the experiments. We also evaluate CAM and SAM using datasets from the UCI repository and the result has been promising.

Specification of Kansei Patterns in an Adaptive Perceptual Space

In this paper, we apply the algorithms to facilitate learning to kansei modeling and experimentally investigate constructed kansei model itself. We introduce using a vector space as a scheme of the mental representation and place still images in the perceptual space by generating perceptual features. Furthermore we propose a method to manipulate the perceptual data by optimizing modeling parameters based on the kansei scale. After this adaptation we compare the similarity between the kansei clusters using their distance in the space to evaluate if the adapting perceptual space is appropriate for ones kansei. We have conducted preliminary experiments utilizing image data of TV commercials and briefly evaluated the mental space constructed by our method through the kansei questionnaire.