新居 学

In order to improve the nursing-care quality, "Web based Nursing-care Quality Improvement System" have been proposed and operating continuously. In the proposed system, for evaluating actual nursing-care process, freestyle Japanese texts which are called "nursing-care texts" are collected through the Internet in Japan. The nursing-care experts can evaluate actual nursing-care process and recommend some improvements to nurses by reading the collected nursing-care texts carefully. Since the number of nursing-care experts who can evaluate the nursing-care texts is a few, it is hard to do the above mentioned evaluation process for a large number of nurses. To assist nursing-care experts in evaluating the nursing-care texts, a computer aided nursing-care text classification system has been developed. In this paper, we propose a method to improve the classification performance of the computer aided nursing-care text classification system. Dependency relation between terms is extracted from the nursing-care text and used the dependency as a feature value which represents characteristics of the nursing-care text.

In order to improve the nursing-care quality, freestyle Japanese texts which are called ''nursing-care texts" are collected through the Internet in Japan. T he nursing-care experts can evaluate actual nursing-care and recommend some improvements to nurses by reading the collected nursing-care texts carefully. Since the number of nursing-care experts who can evaluate the nursing-care texts is a few, it is hard to do the above mentioned evaluation process for a large number of nurses. To assist nursing-care experts in evaluating the nursing-care texts, a computer aided nursing-care text classification system has been developed. In this paper, we propose a method to improve the classification performance of the computer aided nursing-care text classification system. Conceptual fuzzy sets are constructed from the collected nursing-care texts and used to make feature vectors in our proposed method.

This paper presents an advanced method which measures the real-time transmission delay time about all frames from video camera. This camera takes a picture of a timer display which steadily shows shutter-closing time of each frame. Additionally, two synchronous on-off markers have been built in that timer. These components, timer display and on-off markers run using frame synchronous timing of the camera. Two trial equipments succeed to measure the detailed characteristics of delay along the real-time SD-TV transmitted by the H.264 technology or by the DVTS technology with a resolution of 0.1ms. These experiments suggest that the synchronous on-off marker method is useful.

A monitoring system based on multiple microelectromechanical systems (MEMS) has been developed to maintain human healthcare. Using such a MEMS based monitoring system, several kinds of numerical data from several types of sensors can be measured. Our goal is to develop a intelligent monitoring system with small size. In order to microminiaturize the monitoring system, we need to minimize the power consumption. Therefore, we have to keep our intelligent system simple. We propose a behavior estimation method which consists of a SVM and a fuzzy rule based system to estimate the subject's behavior. Our proposed method consists of two steps of abstraction. First, action primitives are defined. A SVM based classification system is trained using sample numerical data of action primitives. Then, the SVM based system classifies a part of numerical data into one of action primitives. Therefore, whole numerical data are expressed by a sequence of action primitives. Next, a fuzzy rule which maps a sequence of actions onto a behavior is defined for each behavior. In the second-step abstraction, each action sequence is expressed as a behavior by using the defined fuzzy rules. From the results of the abstraction, we can estimate the subject's state.

To record daily activity for well-maintained human health care, a monitoring system based on multiple microelectromechanical systems (MEMS) has been developed. Several kinds of numerical data of subject's activity can be stored using the MEMS based monitoring system. When we use subject's activity on a single day, a huge volume of data is obtained and recorded. In order to estimate the subject's behavior from such a huge volume data, we propose a behavior estimation method which consists of a SVM and a fuzzy rule based system. Our proposed method consists of two steps of abstraction. First, action primitives are defined and a SVM based classification system is generated from sample numerical data of action primitives or human knowledge. Then, the SVM classifies a part of numerical data into each action primitive. Therefore, numerical data are expressed by a sequence of action primitives. Next, a fuzzy rule which maps a sequence of actions onto a behavior is defined by human user for each behavior. In the second-step abstraction, each action sequence is expressed as a behavior by using the defined fuzzy rules. From the results of the abstraction, we can estimate the subject's state. © 2011 IEEE.

When we use a MEMS based monitoring system for monitoring human behavior, several kinds of numerical data can be measured. In order to develop an intelligent monitoring system with small size, we need to minimize the power consumption. Therefore, we need to develop a simple intelligent system. We have proposed a behavior estimation method which consists of a fuzzy rule based system to estimate the subject's behavior. In the proposed method, two steps of abstraction for numerical data extract some human behaviors from numerical data. The proposed method can estimate only human behavior such as "walking" or "sitting" etc. In this paper, we propose an estimation method that can estimate both human state and behavior. Our proposed method can estimate human state using both heart rate from ECG and human behavior obtained from acceleration sensors. From the results of our proposed method, we show the effectiveness of our proposed method. © 2011 IEEE.

A wearable health monitoring system and its applications for long term monitoring are presented in this paper. The system, called a "button system," is attached over the chest for monitoring electrocardiogram (ECG), heart rate (HR), 3 axis acceleration, and temperature, as well as system battery voltage. The data is then sent to the host computer via a wireless transmitter. The button system is composed of three round substrates connected together. The substrate is 24 mm in diameter and 5-10 mm thick, making the system applicable for use in daily life. The several various types of sensors, microcontroller, programmable gain amplifier (PGA) and Bluetooth wireless transmitter are embedded in system. The Micro Processing Unit (MPU) samples ECG and acceleration at 125 Hz to calculate HR for transmission. The battery voltage and temperature are sampled at 0.2 Hz due to the slow variability. The current consumption during wireless telecommunications is about 40 mA, so it is possible to use it continuously for about two hours with Li-ion button battery (75 mAh). In this study, we measured data while sleeping (for about 6 hours) with a large capacity battery (2000 mAh) instead of Li-ion battery. Heart rate variability (HRV) was then used as a quantitative marker of automatic nervous system activity, and the temporal variation of HRV was estimated. © 2011 IEEE.

In this paper, for improving performance of the nursing-care text classification, we introduce a mechanism of retrieving terms from Web. Every year, the nursing-care texts are collected by using Web application to improve nursing-care quality in Japan. The collected nursing-care texts are decomposed into morphemes (i.e., terms), and then terms are stored as a term list. Each text is represented as a feature vector by using the term list and classified using a SVM based classification system. The training data sets for constructing SVM based classification system are different from the evaluation data sets. That is, there are differences between the term lists of the nursing-care texts because the nursing-care texts are collected and evaluated every year. To cover this difference, we introduce a mechanism of retrieving terms from Web. A new term which appeared in the evaluation data sets is used as a query of a search engine. The terms in the term list are also used as queries. Terms are represented by the search results, and then are compared with each other. We use the most similar term in the term list as an alternative of the new term. From experimental results, we show effectiveness of our proposed method.

In systems such as chemical plants or circulatory systems, failure of piping, sensors or valves causes serious problems. These failures can be avoided by the increase in sensors and operators for condition monitoring. However, since adding sensors and operators leads to an increase in cost, it is difficult to realize. In this paper, a technique of diagnosing target systems based on a fuzzy nonlinear regression is proposed by using a fuzzified neural network that is trained with time-series data with reliability grades. Our proposed technique uses numerical data recorded by the existing monitoring system. Reliability grades are beforehand given to the recorded data by domain experts. The state of a target system is determined based on the fuzzy output from the trained fuzzified neural network. Our proposed technique makes us determine easily the state of the target systems. Our proposed technique is flexibly applicable to various types of systems by considering some parameters for failure determination of target systems.

In this paper, we propose two term selection methods for classifying nursing-care texts. In a term selection method based on GA, two objectives which are maximizing correctly classified texts and minimizing selected terms are optimized. The weighted sum of these two objectives was used as the evaluation function. Therefore, GA-based term selection is performed aiming at the improvement in classification performance on testing sets. In a NSGA-II based term selection method, non-dominated solutions are found. As the result, we can have a set of pareto-optimal solutions. These solutions are helpful to analyze classification results from the viewpoint of terms. From experimental results, we show effectiveness of our proposed term selection methods. © 2011, TSI® Press Printed in the USA. All rights reserved.

In this paper, for improving classification performance of a term selection based on GA, we modify its evaluation function and mutation operation. In the term selection based on GA, two objectives which are maximizing correctly classified texts and minimizing selected terms are optimized. The weighted sum of these two objectives was used as the evaluation function. Therefore, GA-based term selection is performed aiming at the improvement in classification performance on testing sets. This causes the performance deterioration over completely unseen texts. This is because terms are deleted excessively even when the terms have important role for the classification. First, we use NSGA-II for finding non-dominated solutions. As the result, we can have a set of pareto-optimal solutions. Each individual is evaluated by using SVM with -fold cross validation. In this paper, we also modify the mutation operation. The modified mutation operation uses the statistic information of each term as the mutation probability. From numerical simulation results, we show effectiveness of our modification.

In this paper, classification performance of a term selection based on GA is analyzed. In the term selection based on GA, two objectives which are maximizing correctly classified texts and minimizing selected terms are optimized. An objective function based on the classification per-formance of the SVM with 10-fold cross validation is used for evaluating each individual in GA. Therefore, GA-based term selection is performed aiming at the improvement in classification per-formance on testing text sets. This causes the performance deterioration over unseen texts in actual use by GA-based term selection because terms are deleted excessively even when such terms have important role for the classification. In this paper, relation between the terms deleted by the term se-lection based on GA and the terms which appears in unseen texts is clarified by numerical simulation results. © 2010 TSI Press.

People often want to know the names of the objects that they can explain but don't know the names. It is, however, difficult to find such object names using conventional Web search engines. So, we propose a new method for finding the object name from the descriptions given by a user. This method consists of two phases, the extraction phase and the validation phase. In the extraction phase, candidate words are extracted by conducting a Web search using a combination of the queries generated from the user's descriptions. In the validation phrase, each candidate word is validated through a Web search using the candidate word. We rank the candidate words based on the user's description. We evaluated our algorithm by performing several tasks to find the object names from questions in Q&amp A sites. We also compared it with the methods using queries consisting of all the words in the description and queries consisting of user-selected and user-generated words. The precision by our algorithm was higher than the precision by the other methods. ©2010 IEEE.

In order to maintain human health care, it is important to record daily activity. For recording daily human activity, monitoring system which consists of multiple microelectromechanical systems (MEMS) has been developed. Using the MEMS based monitoring system, numerical data of subject's activity can be stored into a database. For example, when subject's activity on a single day is recorded, a huge volume of data is saved. To estimate the subject's activity condition from such a huge volume data, a fuzzy rule based approach is used in our study. Our proposed method consists of two steps of abstraction. First, action primitives are defined. In the first-step abstraction, sensor data is expressed as a sequence of actions by using the defined action primitives. Next, a fuzzy rule which maps a sequence of actions to a behavior is defined for each behavior. In the second-step abstraction, each sequence of actions is expressed as a behavior. From the results of abstraction, we can estimate the subject's state. © 2010 IEEE.

We define the peculiarity of text as a metric of information credibility. Higher peculiarity means lower credibility. We extract the theme word and the characteristic words from text and check whether there is a subject-description relation between them. The peculiarity is defined using the ratio of the subject-description relation between a theme word and characteristic words. We evaluate the extent to which peculiarity can be used to judge by classifying text from Wikipedia and Uncyclopedia in terms of the peculiarity.

In circulatory systems or systems like chemical plants, failure of piping, sensors or valves causes serious problems. These failures can be prevented by the increase in sensors and operators for condition monitoring. However, since the increase in cost is required by adding sensors and operators, it is not easy to realize. In this paper, a technique of diagnosing target systems is proposed by using a fuzzified neural network which is trained with time-series data with reliability grades recorded by the sensor system which has already existed. Reliability grades are beforehand given to the recorded data by domain experts. The state of a target system is determined based on the fuzzy output value from the trained fuzzified neural network. Our proposed technique makes us determine easily the state of the target systems. Our proposed technique is flexibly applicable to various types of systems by considering some parameters for failure determination of target systems.

In order to reduce evaluation workloads for nursingcare experts, we have proposed a Support Vector Machine (SVM) based classification system. In this paper, for improving the classification performance, we propose a Genetic Algorithm (GA) based attribute selection method. First, we extract nouns and verbs from nursing-care texts by using of the morphological analysis software and store the extracted terms into a "term list." Next, some combinations of terms in the term list are selected by a GA with two objectives (1) maximizing the number of correctly classified texts and (2) minimizing the number of selected terms. Then, we classify the nursing-care texts with these selected terms by using of a SVM-based classification system. From computer simulations, we show the effectiveness of a GA-based attribute selection method for classifying the nursing-care texts.

This paper presents a method which measures the transmission delay time of real-time scene from the video camera onto the distant screen. The camera takes a picture of a timer with on-off marker and sends it to the screen. Another timer ticking away the same time is placed nearby that screen. When the image of on-off marker has been drawn on the screen, this timer catches its light and shows that time. Then, we can read the difference of the two timers as the total delay. A trial equipment succeeds to measure it along the real-time HD-TV displayed on Organic EL and also H.264 on LCD with a resolution of 0.1ms. The experimental values agree with the analytical characteristics.

We propose a method for searching for comprehensible how-to information on the Web. In our how-to information search, we use lightweight analysis of Web pages to extract how-to information from Web pages obtained by conventional Web search engines and rank them according to their easily-viewable-degree. In the extraction process, we focus on expressions in Web page text blocks that describe procedures. In the ranking process, we focus on images, the effect of letter string and the length of the how-to information. Copyright 2009 ACM.

Recently, we have been able to often compare two objects using search engines. However, we often browse high ranked Web pages by search engines, which may give biased information. We propose a method for searching Web pages where two objects are compared using a search engine, extracting comparison points from those Web pages, and showing these points to users. Comparison points are keywords for comparing objects. The proposed method can be used to extract points for efficient comparison by using comparison expressions such as "Liquid Crystal TVs are better ..." and "... than Plasma TVs.", etc. Copyright 2009 ACM.

The nursing care quality improvement is very important for our life. Currently, nursing-care freestyle texts (nursing-care data) are collected from many hospitals in Japan by using Web applications. The collected nursing-care data are stored into the database. To evaluate nursing-care data, we have already proposed a fuzzy classification system [1], a neural network based system [2], a support vector machine (SVM) based classification system [3], [4]. Then, in order to improve the classification performance, we have proposed a genetic algorithm (GA) based feature selection method [5] for generating numerical data from collected nursing-care texts. In this paper, we propose a fuzzy rule extraction method from the nursing-care text data. First, features of nursing-care texts are selected by a genetic algorithm based feature selection method. Next, numerical training data are generated by using selected features. Then we train neural networks using generated training data. Finally, fuzzy if-then rules are extracted from the trained neural networks by the parallelized rule extraction method [6], [7]. From computer simulation results, we show the effectiveness of our proposed method.

医療の現場においてその質を維持・向上していくことは非常に重要なことである．近年，ツールの汎用化と医療現場における実用性を考慮して電算化が図られ，インターネットを介してデータ収集を行い，大量のデータを蓄積するシステム（Web版看護ケアの質評価総合システム）が開発されている．このような取組みは他に例がないため，看護ケアの質向上のために多くの病院の参加が望まれる．しかし現在は，数名の専門家から構成される看護ケア研究班が収集された自由記述回答を実際に読み，記載内容から実施された看護行為を判断して評価しているため，参加病院数を増やすことが難しい．本研究では，Web版看護ケアの質評価総合システムを広く利用してもらうために，自由記述回答の自動分類に機械学習アルゴリズムを用いたテキスト分類システムを構築する．具体的には，あらかじめ専門家により分類されている自由記述回答を用いて，文章の特徴を抽出して数値データ化し，これを教師データとしてサポートベクターマシンにより分類を行う．本論文では，文章の長さや，使用されている単語や表現に関する評価者の知見を反映した特徴ベクトル生成法を提案する．数値実験結果から，提案手法により未知データに対して良好な分類性能が得られることを示す．本研究の成果により多数の看護師からの回答を評価できるようになり，Web版看護ケアの質評価総合システムの利用による看護ケアの質向上が期待できる．

The nursing care quality improvement is very important in the medical field. Currently, nursing-care freestyle texts (nursing-care data) are collected from many hospitals in Japan by using Web applications and stored into the database. Some nursing-care experts evaluate the collected data to improve nursing care quality. For evaluating the nursing-care data, experts need to read all freestyle texts carefully and then classified them into four classes. However, it is a very hard task for each expert to evaluate the data because of huge number of nursing-care data in the database. In order to reduce workloads evaluating nursing-care data, we have proposed a support vector machine (SVM) based classification system. In this paper, to improve the classification performance, we propose a feature extraction method for generating numerical data from collected nursing-care texts. In our proposed method, the frequency in use of a term in the term list is used for selecting features which contribute to the classification. And then, the nursing-care numerical data are classified by the SVM based classification system. From computer simulation results, we show the effectiveness of our proposed method.

The nursing care quality improvement is very important in. the medical field. Currently, nursing-care freestyle texts (nursing-care data) are collected from many hospitals in Japan by using Web applications. Some nursing-care experts evaluate the collected data to improve nursing care quality. For evaluating the nursing-care data, experts need to read all freestyle texts carefully. However, it is a hard task for an expert to evaluate the data because of huge number of nursing-care data in the database. In. order to reduce workloads evaluating nursing-care data, we propose a support vector machine(SVM) based classification system.

In this paper we examine the performance of the evolutionary algorithm for a dynamic environment where the opponent team changes during the evolution of strategies. In the dynamic environment, the dash power of opponent players in ball intercept behavior is adjusted. We consider two adjustment modes of the dash power in our experiments. In one mode the dash power of the opponent players is gradually increased over generation of the evolutionary algorithm. The other mode monitors the performance of the evolved team strategies to increase or decrease the dash power of opponent players accordingly. We compare the evolution of team strategies between the two modes and discuss for the future extension to the current experimental settings. We also discuss the possibility of knowledge extraction from the obtained team strategies by the evolutionary algorithm. ©ISAROB 2007.

Nursing-care data in this paper are Japanese texts written by nurses which consist of answers for questions about nursing-care. The nursing-care data are collected via WWW application from many hospitals in Japan. The collected data are stored into the database. The nursing-care experts evaluate the collected data to improve nursing-care quality. Currently, the collected data are evaluated by experts reading all texts carefully. It is difficult, however, for experts to evaluate the data because there are huge number of nursing-care data in the database. In this paper, to reduce workloads for the evaluation of nursing-care data, neural networks are used for classifying nursing-care data instead of fuzzy classification system. We use standard three-layer feedforward neural networks with back-propagation type learning. First, we extract attribute values (i.e., training data) from texts written by nurses. And then, we train a neural network using the training data. From computer simulations, we show the effectiveness of our proposed system using the leaving-one out method.

This paper proposes an evolutionary method for acquiring team strategies of RoboCup soccer agents. The action of an agent in a subspace is specified by a set of action rules. The antecedent part of action rules includes the position of the agent and the distance to the nearest opponent. The consequent part indicates the action that the agent takes when the antecedent part of the action rule is satisfied. The action of each agent is encoded into an integer string that represents the action rules. A chromosome is the concatenated string of integer strings for all agents. We employ an ES-type generation update scheme after producing new integer strings by using crossover and mutation. Through computer simulations, we show the effectiveness of the proposed method.

This paper proposes an action rule discovery technique from simulated RoboCup soccer logs. Our proposed technique analyzes log files of past games and generates effective action rules for agents. From simulation results, we show the effectiveness of our action rule discovery technique.

In this paper we improve the performance of an evolutionary method for obtaining team strategies in simulated robot soccer. In the previous method each team strategy was evaluated based on the goals and the goals against of a single game. It is possible for a good team strategy to be eliminated from the population in the evolutionary method as there is a high degree of uncertainty in the simulated soccer field. In order to tackle the problem of uncertainty, we propose a robust evaluation method using match history. The performance of team strategies in the proposed method is measured by the average goals and average goals against. Through a series of computer simulations, we show the effectiveness of our robust evaluation method.

In this paper we improve the performance of an evolutionary method for obtaining team strategies in a simulated robot soccer domain. In the previous method each team strategy was evaluated based on the goals and the goals against of a single game. It is possible for a good team strategy to be eliminated from the population in the evolutionary method as there is a high degree of uncertainty in the simulated soccer game. In order to tackle the problem of uncertainty, we propose a robust evaluation method using match history. The performance of team strategies in the proposed method is measured by the average goals and average goals against. Through a series of computational experiments, we show the effectiveness of our robust evaluation method.

本研究の目的は，ファジィ識別システムを並列実装することによる処理の高速化である．ファジィ識別システムは，ファジィ If-Then ルールの集合で構成されており，学習用データの次元数が大きい場合にはルール数の爆発が起こり処理時間が膨大となる．さらに，学習用データ集合のサイズが大きい場合には性能評価にも計算コストがかかる．そこで，本研究では，ルール生成処理とパターン識別処理を並列化することにより計算時間の短縮を図る．まず，従来の単一計算機による実装での計算時間と学習用データの次元数，学習用データ集合のサイズとの関係を実験的に調査する．次に，並列実装したファジィ識別システムの計算時間を調査し，並列化することによりファジィ識別システムの処理時間が短縮されることを示す．

When one application needs results calculated by another application simulating different phenomena to simulate some phenomena, users must write communication codes to exchange data between two applications. Because users require to program codes as simple as possible in general, writing communication codes should be more easier. It is, however, complicated for users to implement these codes using the existing method. In this paper, we have proposed the Distributed Computing Collaboration Protocol as a simple user interface for communication between application programs.

In the parallelized rule extraction method, it is very hard to predict the number of extracted rules on each processor. The difference of the number of extracted rules on each processor will cause the decline of the efficiency in parallel computing. Our system can reduce the difference of the number of extracted rules by using of an autonomous load balancing.

The rule extraction method, which extracts fuzzy if-then rules from trained neural networks, needs to input all combinations of antecedent fuzzy sets. For high-dimensional problems, it is very hard to input all combinations to neural networks because the number of such combinations is exponentially increased. Our attribute selection method can reduce the number of input combinations.

In this paper, we propose a parallelized rule extraction method from trained neural networks for high-dimensional pattern classification problems. In the rule extraction method, we have to examine all combinations of antecedent fuzzy sets for extracting fuzzy rules. For high-dimensional problems, the number of possible combinations is increased exponentially. To address this difficulty, we parallelize the rule extraction method.

The main aim of this paper is to clearly show how fuzzified neural networks are trained by back-propagation-type learning algorithms for approximately realizing fuzzy if-then rules. Our fuzzified neural network is a three-layer feedforward neural network where connection weights are fuzzy numbers. A set of fuzzy if-then rules is used as training data for the learning of our fuzzified neural network. That is, inputs and targets are linguistic values such as "small" and "large". In this paper, we first demonstrate that the fuzziness in training data propagates backward in our fuzzified neural network. Next we examine the ability of our fuzzified neural network to approximately realize fuzzy if-then rules. In computer simulations, we compare four types of connection weights (i.e., real numbers, symmetric triangular fuzzy numbers, asymmetric triangular fuzzy numbers, and asymmetric trapezoidal fuzzy numbers) in terms of the fitting ability to training data and the computation time. We also examine a partially fuzzified neural network. In our partially fuzzified neural network, connection weights and biases to output units are fuzzy numbers while those to hidden units are real numbers. Simulation results show that such a partially fuzzified neural network is a good hybrid architecture of fully fuzzified neural networks and neural networks with non-fuzzy connection weights. (C) 2001 Elsevier Science B.V. All rights reserved.

In this paper, first we explain several versions of fuzzy regression methods based on linear fuzzy models with symmetric triangular fuzzy coefficients. Next we point out some limitations of such fuzzy regression methods. Then we extend the symmetric triangular fuzzy coefficients to asymmetric triangular and trapezoidal Fuzzy numbers. We show that the limitations of the fuzzy regression methods with the symmetric triangular fuzzy coefficients are remedied by such extension. Several formulations of linear programming problems are proposed for determining asymmetric fuzzy coefficients from numerical data. Finally, we show how fuzzified neural networks can be utilized as nonlinear fuzzy models in fuzzy regression. In the fuzzified neural networks, asymmetric fuzzy numbers are used as connection weights. The fuzzy connection weights of the fuzzified neural networks correspond to the fuzzy coefficients of the linear fuzzy models. Nonlinear fuzzy regression based on the fuzzified neural networks is illustrated by computer simulations where Type I and Type II membership functions are determined from numerical data. (C) 2001 Elsevier Science B.V. All rights reserved.

本論文では, 遺伝的アルゴリズムを用いてパターン選択と特徴選択を行うことにより最近傍識別器を設計する場合での適応度関数の定義について議論する.数値実験では, 最近傍識別器の汎化性能と適応度関数の定義方法との関係を調べ, 遺伝的アルゴリズムを用いたパターン選択と特徴選択により汎化性能が向上することを示す.

We examine the effect of instance and feature selection on the generalization ability of trained neural networks for pattern classification problems. Before the learning of neural networks, a genetic-algorithm-based instance and feature selection method is applied for reducing the size of training data. Nearest neighbor classification is used for evaluating the classification ability of subsets of training data in instance and feature selection. Neural networks are trained by the selected subset (i.e., reduced training data). In this paper, we first explain our GA-based instance and feature selection method. Then we examine the effect of instance and feature selection on the generalization ability of trained neural networks through computer simulations on various artificial and real-world pattern classification problems.

This paper discusses the design of classification systems when we have two kinds of information: numerical data and linguistic knowledge. Numerical data are given as a set of labeled samples (i.e., training patterns), which are usually used for designing classification systems in various pattern classification techniques. Linguistic knowledge is a set of fuzzy if-then rules, which is not usually utilized in non-fuzzy pattern classification techniques. In this paper, it is implicitly assumed that either kind of information is not enough for designing classification systems with high classification performance. Thus our task is to design a classification system by simultaneously utilizing these two kinds of information. In this paper, we illustrate two approaches to the design of classification systems from numerical data and linguistic knowledge. One is a fuzzy-rule-based approach where numerical data are used for generating fuzzy if-then rules. The other is a neural-network-based approach where linguistic knowledge as well as numerical data are used for training neural networks. First we discuss the extraction of fuzzy if-then rules directly from numerical data. We also describe the fuzzy rule extraction from neural networks that have already been trained using numerical data. Next we discuss the learning of neural networks from numerical data and linguistic knowledge. In the learning, fuzzy if-then rules and training patterns are handled in a common framework. Finally we examine the performance of these approaches to the design of classification systems from numerical data and linguistic knowledge through computer simulations.

This paper discusses various techniques for soft decision making by neural networks. Decision making problems are described as choosing an action from possible alternatives using available information. In the context of soft decision making, a single action is not always chosen. When it is difficult to choose a single action based on available information, the decision is withheld or a set of promising actions is presented to human users. The ability to handle uncertain information is also required in soft decision making. In this paper. we handle decision making as a classification problem where an input pattern is classified as one of given classes. Class labels in the classification problem correspond to alternative actions in decision making. In this paper, neural networks are used as classification systems, which eventually could be implemented as a part of decision making systems. First we focus on soft decision making by trained neural networks. We assume that the learning of a neural network has already been completed. When a new pattern cannot be classified as a single class with high certainty by the trained neural network, the classification of such a new pattern is rejected. After briefly describing rejection methods based on crisp outputs from the trained neural network, we propose an interval-arithmetic-based rejection method with interval input vectors, and extend it to the case of fuzzy input vectors. Next we describe the learning of neural networks for possibility analysis. The aim of possibility analysis is to present a set of possible classes of a new pattern to human users. Then we describe the learning of neural networks from training patterns with uncertainty. Such training patterns are denoted by interval vectors and fuzzy vectors. Finally we examine the performance of various soft decision making methods described in this paper by computer simulations on commonly used data sets in the literature. (C) 2000 Elsevier Science B.V. All rights reserved.

We have already shown that the relation between neural networks and linguistic knowledge is bidirectional for pattern classification problems. That is, neural networks are trained by given linguistic rules, and linguistic rules are extracted from trained neural networks. In this paper, we illustrate the bidirectional relation for function approximation problems. First we show how linguistic rules and numerical data can be simultaneously utilized in the learning of neural networks. In our learning scheme, antecedent and consequent linguistic values are specified by membership functions of fuzzy numbers. Thus each linguistic rule is handled as a fuzzy input-output pair. Next we show how linguistic rules can be extracted from trained neural networks. In our rule extraction method, linguistic values in the antecedent part of each linguistic rule are presented to a trained neural network for determining its consequent part. The corresponding fuzzy output from the trained neural network is calculated by fuzzy arithmetic. The consequent part of the linguistic rule is determining by comparing the fuzzy output with linguistic values. Finally we suggest some extensions of our rule extraction method.

When a fuzzy input vector is presented to a multi-layer feedforward neural network, the corresponding fuzzy output vector is calculated by fuzzy arithmetic. It is well-known that fuzzy arithmetic involves excess fuzziness. In this paper, we employ subdivision methods of interval input vectors for decreasing excess fuzziness included in fuzzy outputs from neural networks: First we examine a simple subdivision method where each level set of a fuzzy input vector is subdivided into many cells with the same size by uniformly subdividing all elements of the level set into multiple intervals. Next we examine a hierarchical subdivision method where each level set is subdivided into many cells with different sizes by iteratively subdividing a single element of a cell into two intervals. Finally Mle modify the hierarchical subdivision method for efficiently decreasing excess fuzziness.

<p>ニューラルネットワークにファジィ数ベクトルが入力された場合, 対応するファジィ数出力ベクトルの計算は, ファジィ演算を用いて行われる.このとき, ファジィ演算が各ユニットで独立に実行されるため, 入力ベクトルに含まれるあいまいさが, 入力層から出力層へ向けた計算の進行とともに増加するという問題点が生じる.本研究では, まず, この問題点を, 拡張原理の局所的な適用と大域的な適用という観点から議論する.局所的な適用では, 各ユニットでのファジィ入出力関係が拡張原理により定義される.一方, 大域的な適用では, ニューラルネットワーク全体のファジィ入出力関係が拡張原理により定義される.次に, 具体的な数値計算例を用いて, 拡張原理の局所的な適用におけるあいまいさの増加がどの程度であるかを示す.さらに, あいまいさの増加を低減するため, レベル集合の細分化手法を示す.最後に, ニューラルネットワークからのファジィルールの抽出に細分化手法を適用し、その有効性を検討する.</p>

We have already proposed a fuzzy arithmetic-based method for extracting linguistic if-then rules from trained neural networks for pattern classification problems with continuous attributes. In our method, antecedent linguistic values of a linguistic if-then rule are presented to a trained neural network as inputs, and the corresponding fuzzy outputs are calculated by fuzzy arithmetic. The consequent class and the grade of certainty are determined based on the calculated fuzzy outputs. Thus the calculation of the fuzzy outputs is very important for the linguistic rule extraction. Because the fuzzy arithmetic is locally applied to the calculation at each unit, the fuzziness of the linguistic input values is usually increased by the feedforward calculation through the neural network. In this paper, we show how such increase of the fuzziness can be reduced by subdividing the level set (i.e., alpha - cut) of each linguistic input value in the calculation of the fuzzy outputs. The effect of such subdivision is illustrated by computer simulations.

Recently, interval-arithmetic-based neural networks have been proposed for handling intervals as inputs of multi-layer feedforward neural networks. This paper demonstrates that interval arithmetic can be utilized for improving the generalization ability of neural networks for pattern classification problems. We examine two approaches, each of which is used in the classification phase of new patterns and in the learning phase of neural networks, respectively. In the first approach, an interval input vector is generated from a new pattern by adding a certain width to its attribute values. In the second approach, neural networks are trained by interval input vectors generated from training patterns. These approaches are illustrated by a two-dimensional pattern classification problem. The effectiveness of these approaches is examined by computer simulations an a commonly used benchmark data set.

The aim of this paper is to clearly demonstrate that the relation between neural networks and linguistic knowledge is bidirectional. First we show how neural networks can be trained by linguistic knowledge, which is represented by a set of fuzzy rules. Next we show how linguistic knowledge can be extracted from neural networks. Then we discuss the design of classification systems when numerical data and Linguistic knowledge are available. Since the relation between neural networks and linguistic knowledge is bidirectional, we can simultaneously utilize these two kinds of information for designing classification systems. For example, neural-network-based classification systems can be trained by numerical data and linguistic knowledge. Fuzzy rule-based classification systems can be designed by linguistic knowledge and fuzzy rules extracted from neural networks. The performance of these classification systems is examined by computer simulations.

In this paper, we propose an approach for improving the generalization ability of multi-layer feedforward neural networks. Our approach is based on the fuzzification of input vectors. In our approach, a neural network is trained by fuzzy input vectors. The aim of such fuzzification in the learning phase is to avoid the overfitting of the neural network. In the classification phase, each new pattern is fuzzified, and the fuzzy input vector is presented to the trained neural network. The classification of each new pattern is performed based on the corresponding fuzzy output vector from the trained neural network. The aim of the fuzzification in the classification phase is to reject the classification of new patterns close to the classification boundary. The introduction of a reject option can decrease the misclassification rate on new patterns. We examine the effectiveness of our approach by computer simulations on real-world pattern classification problems.