One of the methods of utilizing patent publications, which are intellectual property information, is to classify them from their own viewpoint and grasp trends. However, people meet the following issues. (1) The results diversify because each person's perspective on grasping the contents is different.(2) It is difficult to share the accumulated information with others because the results of classification and the diversification of classifications.(3) It takes a lot of man-hours to understand the contents. To support this task, I am developing a Rough Set Theory based categorization system for patents with machine translation. I explain the algorithm of this system and the evaluation result about it in this paper. The proposed system makes decision rules based on Rough Set Theory from categorized patents. Then it categorizes unlabeled patents by matching rules with them. Those rules are based on terms, so the flexible expression is big problem. To solve it, I make decision rules from translated patents by machine translation system. I implemented a prototype system based on this approach and categorizing patent publications in Japanese with experts. My approach shows the accuracy as about 0.5. The performance of my system is almost the same as the performance with Naive Bayes Classifier. I say my approach has a possibility to reduce the load of categorization patents, but the performance is not enough. To improve my approach, I will analyze the experimental results, propose new algorithm and evaluate using actual patents.

When the issue has various aspects, discussion tends to diverge and generate unsatisfying conclusions. In this paper, we propose the new model for narrowing down the issues which is supported by rough set theory and voting process. To test the rationality of the decision of the model, we carried out the simulation and workshops. As a result, it was shown that the decision of the model gets the consent of the majority.

This paper will introduce some examples of soft computing algorithm development and research for FPGA devices. In particular, I would like to explain the contents of "Hardware for Soft Computing and Soft Computing for Hardware" published by Springer. This paper would like to introduce the design and evaluation of FPGA-based hardware for Soft Computing, which has been researched in recent years.

This paper introduces the relationship between traditional feature engineering approaches and recent deep neural network approaches on image recognition tasks. The tasks aim to understand objects and their situations from given images as we can understand flexibly. Feature engineering approaches are based on statistical modeling in natural language processing and enable to process a large-scale image collection. In general, these approaches first describe local image features from a lot of image patches, then the local features are encoded into a global image feature. Here, the encoding phase purposes to represent compact feature vectors while retaining the information. Both describing local image features and encoding them into a global image feature are developed from empirical and theoretical knowledge of researchers and engineers. Recently, deep neural networks have dramatically improved accuracy or error in several tasks and the knowledge has been applied to extend the network models. This paper briefly introduces the relationship between the feature engineering approaches and the recent extended neural network models.

For elucidation of information processing in the brain, it is useful that a living neuronal circuit simpler than the neuronal network in a whole brain. Neurorobot is a fundamental model system of a small brain in which a neuronal circuit is modified by interaction between neurons and outer environment. In this study, a living neuronal network was divided into right- and left-labeled electrodes sets, and we developed the system in which a moving-robot performed an obstacle-avoidance-behavior according to the number of spikes in spontaneous activity after the electrical stimulation to the right- and left-labelled electrodes sets. Characteristics of slow change in the number of electriacal spikes evoked by continuous stimulation were analysed. As a result, the robot was able to avoid the first obstacles, and accompanied continuous stimulation depressed the spontaneous activity, and then the dynamic changes of the activity drastically influenced on the robot behavior. Thus, then the obstacle avoidance of the robot changed to be failure after the first obstacles, suggesting that a cultured neuronal network is enough fixible to modulated by neuronal activity, and that hysteresis should be taken into consideration for the controlof the neurorobot.

In order to understand human flexible cooperative behavior in a group, we have developed a cooperative pattern task requiring switching of intention and behavior. Through the analysis of the behavioral experiment using the task, subjects’ optimal behavior to a shared target in minimal steps and majority voting based action selection are clarified. In this research, agent models including "self- preferred selection" and "estimation of other agents’ intention" are constructed to reveal implicitly shared unobservable strategy under the subjects’ behavior. Moreover, we analyze the subjects’ action selection process by model-based agent simulation and comparison between subjects’ and agents’ behavior.

In this study, we have designed appropriate combination of pre-processing and classifiers to identify human thought contents from measured brain waves. Analysis data include 5 subjects of data set 1 of BCI Competition IV. Subject recalls left and right hand movements. The differences at that time were identified by two classifications by CNN. An electrode layout was set in consideration of the spatial proximity of the electrodes of electroencephalogram measurement. The number of time divisions of the measured data was set as the number of channels of CNN. After that, the subject's movements were identified from the brain wave data given the spatial structure. As a result, the discrimination rate was 58.8%.

Monte Carlo tree search (MCTS) has a problem that in a huge game where there are hundreds of millions of branching factors of a search tree, a large number of hands are not considered at all in the first step. Grouping nodes is one solution to this problem. To investigate the effect of introducing abstract nodes by grouping on the efficiency of MCTS based on the existing research on grouping nodes, we proposed a method to create game trees randomly, and played experimental matches between grouping AI and no grouping AI. As a result of the experiment, it was shown that the efficiency of the MCTS could be greatly improved by grouping when the branching factor of the search tree is large, or by grouping to reduce more branching factor and increase the depth by that amount.

In this paper, we develop the incomplete information game Puyo Puyo AI. We develop a program to acquire board information from the Puyo Puyo play video and estimate the player's action by using deep learning.

In team sports such as soccer, the instruction skills of coach are important for improving a team performance. The coach is required to have a skill of game analysis by observing the entire game. In order to evaluate a skill of the game analysis, it is necessary to analyze the observation technique actually performed by the human expert. However, it is hard to collect such data from games of real human sports because it requires much cost to prepare the games. In this research, we use the RoboCup soccer simulation environment to solve this cost problem, and try to collect and analyze human’s attention information using a gaze tracking device.

RoboCup Soccer Simulation 2D League has been studied on individual ability development such as pass and small team play. In recent years, tactics have attracted attention as a team-wide cooperation. In this paper, team tactical theory called“5-lanes theory”was implemented and evaluated. In experiments, each team before and after implementation of“5-lanes theory”was played against several teams that participated in the world tournament. After that, we evaluated the number of wins and losses, ball possession from the log information of the game. As a result, the team that implemented the“5-lanes theory”showed a significant difference in ball possession for particular teams.

In RoboCup Soccer Simulation 2D League, agents play soccer games on 2D fields. Although there are various tactics for soccer, it is difficult to evaluate players and teams by specific indices due to the diversification and compatibility of tactics. Therefore, it is important to determine the index which evaluates a specific situation and to show the effectiveness. In recent real soccer, an evaluation index called packing rate is used. The packing rate is an index that measures how many opponents have been able to pass in one pass. The higher the packing rate, the better the success of the game. This paper discusses the effectiveness of the packing rate in the 2D league and verifies the usefulness in RoboCup.

The purpose of this research is to study an algorithm to estimate the state more accurately from the information of the 6-axis sensor mounted on the Mini 4WD. The six-axis sensor is a sensor that measures six variables of acceleration of three axes and angular velocity around three axes by MEMS. By performing these integrations, it is possible to estimate the position, velocity and direction angle by inertial navigation calculation. However, due to the inclusion of second-order integrals and the inevitable stationary noise contained in these sensors, divergences and large errors in the estimates are inevitable. For this reason, in this research, we make use of the fact that the Mini 4WD race is performed with a combination of prescribed course parts, and attempt to raise the accuracy of state estimation with this as a standard by creating a course model. For noise-based observation systems, there are Kalman filters and particle filters, but it is difficult to apply directly to the course-constrained mini 4WD system model because it contains nonlinear and piecewise non-differentiable motions. Therefore, we assumed the nonlinear course model to be position-dependent movement switching, performed model calculation, and attempted correction by integrating the results.

We investigated the possibility of collision experiment of automobile by using Mini 4WD equipped with AI function on behalf of the car. Discuss what kind of conversion is possible based on the scale rule of automobile and 1/32 mini 4WD. Initially we will introduce a special battery "MaBeee" which can control running with smartphone and experiment on the possibility of speed control.

While several sensor data are required for Mini 4WD AI to learn, acquiring sensor data is time-consuming and take a very high cost. Thus, in this study, we developed a simulator for Mini 4WD AI, and learn Mini 4WD AI by the Mini 4WD Simulator.

Mini 4WD Machine has to run in high speed but remains quite stable while a racing game． In order to achieve the previous, Mini 4WD should adjust its speed depending on its running course such as a straight course, a curve and a jump ramp． In this paper, we propose a self-position estimation method for our Mini 4WD AI utilizing digital camera images．

There are several types of generalized fuzzy c-means that include size variables as well as cluster covariance variables. We overview these methods and focus on two topics. First, noise clustering technique can be generalized to have new two methods. Second, these fuzzy models are compared with statistical mixture models. As a result we have a new view to regard a fuzzy model of clustering as a new statistical model. Problems to be considered in this regard are described.

This report presents an experimental study for the behavior of the fuzzy classification function for a generalized fuzzy clustering method including the Bezdek-type fuzzy c-means, the entropy-regularized fuzzy c-means, the power regularized fuzzy c-means, and the Tsallis entropy-based fuzzy c-means. As a result, if the fuzzification parameter value with the first term of the objective function is greater than one, the fuzzy classification function value converges to the reciprocal of the cluster number, independent with the fuzzification parameter value with the seconf term of the objective function.

Fuzzy c-regression is an extension of fuzzy c-means to switching regression, which is useful for estimating multiple regression models from mixed databases consisting from some different data groups. This paper proposes a novel method for emphasizing the differences among local regression models with the goal of clarifying the differences among clusters. The characteristics of the proposed method is demon- strated through an application to residential solar electric power analysis considering the mutual relation among solar power with meteorological elements.

Clustering is the data mining method that does not use supervised data. Several clustering methods, such as hard c-means, require the number of clusters in advance. These methods need criterion which estimates the number of clusters such as cluster validity measures. Cluster validity measures are quantitative evaluation criteria based on the geometric shape of cluster partition. Although, Many cluster validity measures have been proposed, cluster validity measures based on the Voronoi diagram that is geometric shape related to clustering is not discussed. In this paper, we proposed cluster validity measures based on Voronoi diagrams. Through experiments, we confirmed that the proposed method can evaluate adequately the artificial data that cannot be evaluated adequately using Xie-Beni’s index.

It is known that k-means and k-medoids clustering are conventional clustering methods. Gen- erally, these methods use the squared L2-norm as a dissimilarity. However, it is difficult to obtain an adequate result by several conventional methods using the squared L2-norm to complex data. It is nec- essary to consider dissimilarity, which considered the structure of the data more precisely. In this paper, we propose JS-divergence based k-medoids, which considers dissimilarity between local distributions. The local distribution is estimated from an object and its neighbor by kernel density estimation(KDE). KDE is a technique to estimate the unknown probability density function, based on a sample of points taken from that distribution. The effectiveness of the proposed method is described by comparing with several conventional clustering methods through numerical experiments. Furthermore, the influence of parameters use in KDE is described. These results showed that the proposed method is effective.

Fuzzy clustering is given by assigning to each data a set of membership degree to clusters. Quality of such a clustering is evaluated by measuring compactness and separateness. In crisp clustering, many clustering indices are proposed. In this article, we try to fuzzify these clustering indices by using membership degrees with higher order exponent and evaluate the effect in optimization problem.

In recent years, research and development have been carried out to characterize communica- tion robots. When robots with personality coexist in human society, it is important to form their own personalities considering the surrounding environment. In this paper, we propose a model in which the communication robot forms its personality from its own temperament and the external environment. As the result, different temperaments or external environments forms the different character of robots. This indicates that the proposed model forms a character in accordance with its own temperament and the external environment.

In this paper, we propose a model for robots to decide their role in their group. Recently, there has been an increasing effort to develop personal robots in human communities. In order for robots to behave naturally in human communities, robots are required to acquire sociality. For obtaining sociality, robots are needed to do collective action. Collective action is established by members of a group having their roles. So, we propose the robot model that the robot can recognize its own role. And we examined whether the robots using the model can recognize their role when they form a group. As a result, the smaller the group, the earlier the robots recognize their role in the group.

Today, VR technology is used in various situations. However commonly used VR systems cannot present enough haptic feedback. Therefore, even if a button is pushed in the virtual space, hands or arms go through the button. In order to solve this problem, we implemented and evaluated a system in which a drone is flown according to the position of a virtual button, and users push the drone to feel haptic feedback. Because the button pushing interaction requires enough strong and rapidly force, the drone has rotors to move machine rapidly and push a user. This rotor is called "forward propulsion rotors" and a drone equipped with these rotors is called a "forward propulsion drone". The system was evaluated in two ways. In the first evaluation, the force provided by forward propulsion drone was measured. As a result, the force provided by forward propulsion drone is enough to present interaction of pushing a button. In the second evaluation, a user evaluation was performed by combining a forward propulsion drone with a VR environment. As a result, users feel an enough repulsion force is provided by forward propulsion rotors.

This paper proposes a discrete-time piecewise multi-linear (PML) system and a PML controller based on input-output feedback linearizations. In this paper an input-output feedback linearization is used to transform each piecewise model into a linear system known as the Brunovsky canonical form. Thus the whole piecewise system can be stabilized by a feedback linearized controller. Examples are shown to confirm the feasibility of our proposals by computer simulation.

In this paper, we introduce a distribution system of synthesized data of Japanese population using Interdisciplinary Large-scale Information Infrastructures in Japan. In order to implement real- scale social simulations for real communities, model designers need information on citizens in the target communities. We have already developed the synthetic algorithm to synthesize the whole population of Japan using nation-wide, prefecture-wide, city-wide, small district-wide, and basic unit block-wide in Japan. We are currently developing a distribution system using computing systems in Osaka University and Hokkaido University. We show the distribution system of synthesized population, and how to utilize the synthesized population.

The determination of distance based toll for urban expressway is formulated as the problem to define toll function corresponding to the traffic conditions. The complicate evaluation process is requited to involve the nonlinear mathematical programing for analyzing the traffic conditions. Therefore, the evaluation indices for urban transport networks is estimated according to the variable traffic demand and toll function with several parameters. In the study, the optimization model of toll function for urban expressway is constructed with deep learning approach to describe the above process. In particular, the toll function determination model is formulated by convolutional neural networks. Finally, the optimization model of toll function for urban expressway is established.

The bicycle transport policy is promoted in local cities of Japan regarding with urban environment and health of citizens. In particular, the bicycle can be used as a single transport mode as well as a feeder service for the other modes. Therefore, the transport policy relating with complex decision process of trip makers should be important. According to above consideration, the nonlinear modal split modes would be formulated by fuzzy reasoning method in the study. It is known that fuzzy reasoning model is more applicable than the traditional disaggregate model (e.g. logit model). As fuzzy reasoning model is created with person trip survey, the evaluation of future bicycle transport policy can be performed. It is concluded that the impact of bicycle transport policy can be estimated quantitatively.

In recent years, traffic surveys using Wi-Fi packet sensors have been conducted, and it has gradually become clear that tourist excursion behavior can be quantitatively grasped, but no application example on remote islands can be seen. In this study, we will investigate the possibility of grasping the actual conditions of island excursions by the same survey method. There is a World Natural Heritage site in the area, and continuous monitoring of visitors is essential from a natural environmental management point of view, but traditional survey methods are still difficult in terms of cost and personnel. In this survey, sensors were installed at major traffic nodes in the Yaeyama Islands to monitor tourism (including inter-island travel). As a result, we clarified excursions between remote islands on the first day of tourists.

The purpose of this study is to model the evaluation structure of residential area in a regional urban area considering the heterogeneity of households. In addition to indicators such as convenience in transportation, living comfort, and safety and security, it is assumed that the familiarity to community and the living near family are also considered as factors to evaluate the living environment. Therefore, we conduct the internet questionnaire survey in a regional urban area, using the subject as a person who is planning to relocate. Using survey data, we analyze the characteristics of the various indicators and their weights for the residential environment. Furthermore, we also analyze the relationship between household characteristics and the weight of various evaluation factors. Based on the result, we model the evaluation structure of the environment in the residence according to the household characteristics with hierarchical Bayesian modeling. As a result, it can be mentioned that household characteristics should be considered to model the evaluation structure of the residential environment.

In recent years, statistical surveys have been used for understanding actual traffic situations. However, it is difficult to discern travel or routes between different facilities within a narrow area, such as the sphere of influence of a train station, on a real-time basis. One type of big traffic data, a Wi-Fi packet sensor, is capable of measuring mobile terminals with effective Wi-Fi within a sphere of about 200m at all times. However, since the Wi-Fi packet sensor does not provide an exhaustive survey, it is necessary to make compensations. It can be considered possible to identify the traffic flow within a station's sphere of influence by combining the data measured by a Wi-Fi packet sensor with big traffic data containing complementary information. The purpose of this study is to devise a method to estimate the traffic flow within the sphere of influence of a railway station by combining multiple types of big traffic data.

Barrier-free building and infrastructure facilities in the city have progressed to some extent, but regional responses have not progressed. Under such circumstances, vehicles and carts capable of moving up and down steps and stairs without human assistance are expected. Although the stair-climbing robot using the deployment wheel proposed by us is suitable for such a use, a high output motor is required because the load at the stair-climbing is excessive. In this research, it is possible to reduce the load on the motor and minimize the rocking by controlling the amount of expansion when moving up and down. The outline of the control method is described, and the effects are verified.

The Ministry of Land, Infrastructure, Transport and Tourism has formulated a routine inspection procedure for road maintenance and repair. On the other hand, much of the social infrastructure in Japan is deteriorated. Therefore, it is necessary to streamline these periodic inspections. In the existing research, it has been proven that deformation and deterioration of road surfaces can be confirmed without dispatching inspectors by using point cloud data and images acquired through mobile mapping systems. However, it takes time to visually find a target feature from within a huge amount of data. In this research, we project point cloud data and generate images. Then, we propose a method to identify road features using projected images by deep learning.

In Japan, where the aging of social infrastructure is advancing, the momentum for working on infrastructure maintenance throughout society is increasing. Focusing on the road field, laser scanners that can measure the three-dimensional shape of the surrounding environment as point cloud data are attracting attention. Since laser scanners can grasp a wide range of three-dimensional shapes in a short time, research in improving the efficiency of road maintenance using measured point cloud data is being promoted. However, point cloud data is a collection of XYZ coordinates, and these points do not contain information on road objects. It is necessary to select point cloud data relating to the necessary road objects in accordance with the intended application in order to achieve efficient operation in practice. Therefore, technology to automatically recognize road objects from out of a large amount of point cloud data is required. In this research, we propose a method for recognizing road objects using point cloud data by machine learning.

Currently, the infrastructure built during the period of high economic growth is aging, and the importance of maintenance is increasing. However, ascertaining the extent of the damage is expensive because visual and manual inspection are mainstream. Existing research using point cloud data and image data has been conducted to solve this problem. However, there are many limitations, such as high measurement costs and filming conditions, so it is difficult to use in daily inspections. In this research, we propose a road support system that can be patrolled regularly at low cost by extracting inspection target features from video data.

In waste incineration plants, a waste quality is very difficult to be identified even using many sensors. Thus, it is an important subject to build a simulation in the presence of uncertainty. An effective modeling technique that combined neural networks with physical systems is proposed for estimating future steam temperatures in a boiler system. The physical systems incorporate available prior knowledge about the process being modeled, while the neural networks compensate errors of unmeasured process that are difficult to model with the physical systems. Experimental results for real plant data show that the combined model is able to interpolate and extrapolate much more accurately and better than a neural network only model. The model requires significantly fewer training examples and is easier to analyze and interpret the trained model.

In this manuscript, a user-friendly design and training application for convolutional neural networks (CNNs) and support vector machines (SVMs) is introduced. Two kinds of binary classification systems using CNNs, SVMs and template matching techniques are designed for classifying input images into OK category or NG one including small defects, then the classification results and the usefulness of the tool are shown.

We developed a system for detecting anomaly in mechanical equipment with neural network using sensor data. Mechanical equipment have many components and it is difficult to collect many anomaly data. In this research, we focused on the valve only and made an autoencoder that reconstructs only normal data with a neural network. We evaluated the autoencoder using data obtained from artificial trouble tests, where control parameters are changed based on the anomaly in the past. As a result, it was possible to discriminate the occurrence of anomaly.

We discuss on developing an autonomous mobile robot for supporting person to move in a building. Recently, techniques of artificial intelligence and object identification have been studied. Self- driving cars and so on have been developed based on the techniques. We try to develop an autonomous mobile robot which can guide people in a building safely, who don’t know there.

We propose safeness of an autonomous mobile robot for supporting person to move in a bilding. Robots in the future are required cooperative operation through networks. Then it is possible for functions of the robot to be danger from the security point of view. We propose a mutual authentication method for confirming robots are safe.

We propose a network interface of an autonomus mibile robot for supporting person to move in a bilding, where we assume Operating System as Robot OS (ROS). ROS has a function for communication in http protocol. It is hard for the robot to measure its position by GPS because it is in the building. The robot can avoid barriers effectively by communicating various spots in the building. We develop network interface by JacvaScript and TypeScript and estimate the ability.

This paper deals with properties of views that attract visual attention. An experiment in which participants wore an eye tracking system and walked in a tourist site was carried out. The targets of the gaze which attracted the visual attention were examined. A comparison between Japanese and foreign people is also conducted. As a result, it is revealed that both Japanese and foreign people are attracted to objects that are clearly visible, but only foreign people are more attracted to objects where there is spatial changes.

An intellectual skill is needed to read employment information during job hunting. An experiment was conducted in which participants read employment article and their eye movements were recorded with eye tracker. The authors analyzed the elements where they paid much attention in the article and order in which they obtained information. The effect of job hunting experience is also investigated. As a result, it is revealed that there is strategy of reading them by experience of job hunting.

Currently, real-time parking space detection using wide-angle surveillance camera is already in practical use. However, since the current system detects cars using still images, small cars hidden behind larger cars cannot be detected correctly. This research aims to solve this problem by applying motion detection and image recognition to streaming video data to improve car detection accuracy, without adding surveillance cameras to physically decrease blind spots.

Although the accuracy and processing capabilities of machine translators have increased greatly in recent years, the question remains as to whether human translators can ever be completely replaced by software without an appropriate model of human feelings. Toward that objective, we propose several different factors that machine translation may use to present results with higher accuracy. The factors are classified as target audience, text genres, context, emotion intensity, and emotion types. We apply three methods in a survey to elicit the respondents’ actual feelings about a given set of translations by varying specific word choices in the translation samples. Analysis of Pearson correlation ratio shows that context and emotion intensity offer potentially useful information for word choice in translation. These results suggest that the result of word choice can be correlated to the proposed factors and varies depending on text genres.

For effectively and efficiently screening potential sleep-disordered patients, it is desired to establish sim- plified and easy-to-use techniques to measure the sleep state in a sufficient accuracy without elaborate facilities. Pressure sensor mats are a promising candidate but methods to estimate the vital signs of sub- jects from their sensing data are to be devised. For non-invasive sleep disorders screening, Generalized Morphological Component Analysis (GMCA), which is a kind of blind source separation methods, is ap- plied to estimate the respiratory effort curve and its spatial distribution from time series of sleeping body pressure distribution. This study explores respiratory motion features that are effective for detecting and discriminating respiratory events such as apnea and hypopnea by examining the results of applying GMCA to the measurement data of the pressure sensor mat collected in polysomnography tests.

It is useful for patients and therapists to undergo rehabilitation using techniques such as Virtual Reality(VR). This study aims at the improvement of arm function by adding sense of touch information to sight information to be provided in VR. As a method, this study obtained feedback information by sense of touch using Leap Motion Controller, Unity, KKHMF UNO R3 and vibration motors. The vibration motors are attached to each real finger-tip to obtain the sense information of touch of the finger in projected virtual space. When the system recognized that subjects contact with an object in the virtual space, the system made the vibration motors activate. A comparison experiments with and without the vibration motor was to test it and investigated how feedback information affected it.

In information-oriented society, opportunities of using the smartphone for a long time increase, especially for young person. The long-time use of smartphone makes abnormal posture to the young person, and it can cause to make many symptoms. We have developed a posture measurement device calculating the angle of the upper body using gyro sensors. The posture measurement device judges a malposture and can evaluate the risk of the upper body posture. In the experiment, we examined the posuture while using the smartphone for 5 minutes.

Triangular Fibro-Cartilage Complex (TFCC) injures occur due to falling, traffic accident, abuse of wrist at work and log time sport. There are imaging and palpation for diagnosing TFCC tear. Imaging can confirm ligament tear, however it cannot measure bone movement. On the other hand, there are ballottement test and piano key sign in palpation methods, but these methods largely depend on doctor’s sense and experience. Therefore, it is impossible to diagnose quantitatively. This study has developed a measurement system to assess instability of ulna toward palmer side and wrist pressure quantitively in real time by using 3D electromagnetic sensor and pressure sensor.