We have proposed the framework of Rough Sets Non-deterministic Information Analysis (RNIA) and are developing the execution environment. In this paper, we introduce the new execution environment in Python and show some examples. Then, we consider a framework of machine learning based on RNIA. In RNIA, we investigated certain rule generation from tables with missing values. We can recover a value for some missing values by using certain rules and sequentially revise this table to a table without missing values. We term this functionality Machine Learning by Rule Generation from incomplete information.

Weak Kleene (WK) and Paraconsistent Weak Kleene (PWK) are 3-valued logics that take a third truth value other than true or false to represent “meaninglessness”. WK and PWK are also called infectious logic because if a propositional variable in a formula is interpreted as meaningless, then the interpretation spreads throughout the formula. This study aims to clarify the relationship between WK and PWK logic systems and rough sets from algebraic semantics. In particular, we investigate the possibility to show the construction of an algebraic model of WK and PWK by Plonka sums using rough set logic.

In this paper, we discuss an improved method to quickly calculate relative reducts after some object was deleted from the decision table. In general, when a decision table that has already created relative reducts is updated, it is necessary to calculate all relative reducts again. However, recalculation is very computationally intensive and takes a lot of processing time. We proposed a method to quickly calculate relative reducts after some object was deleted from the decision table. However, the problem with this method is that it cannot be used for decision table with inconsistent data. Therefore, we propose a new method that can be used even with inconsistent data.

In this paper, we investigate a method for inducing k-anonymous rules from a data table about a binary classification toward the rule publication. Although imprecise rules work well for inducing k-anonymous rules from data tables about multiclass classifications, they cannot be applied directly to binary classification data because imprecise rules for binary classification become trivial. The method of subdividing classes has been investigated. However, the method requires a great computational effort caused by plentifully generated subclasses. We proposed an efficient method for inducing k-anonymous rules from a data table about a binary classification. In the proposed method, after the application of a usual induction method, we concentrate on k-anonymous rule induction covering instances which have not yet covered by any k-anonymous rules. The effectiveness of the proposed method is examined by numerical experiments.

The number of traffic accidents tends to increase during twilight compared to daytime. Until now, many studies have been conducted from the perspective of car drivers. In this study, we analyzed the difference in visual behavior of pedestrians between daytime and twilight using objective data obtained from an eye tracking device. As a result, in twilight, number of long fixations increase, short fixations spread, and long fixations converge space comparing with in daytime.

Glass-type eye tracking device records gaze direction of participants in the visual field. In order to integrate and compare the gaze data, it is necessary to transform the two-dimensional visual field coordinate system data into the three-dimensional common coordinate system data. This transformation requires visual distance data. Self-localization method used in robotics to estimate one's own position on a map can be used to estimate the visual distance in theory, but it is not possible in practice. Because, in eye tracking application, users have to match points in the visual field to those on map manually one by one. The authors developed an interface to support this task and made it possible to estimate the visual distance practically.

The basic means of information gathering for international students is the Internet. Therefore, a university's website must be appealing and informative. In this paper, we focus on attention allocation during the information-seeking and conduct a comparative experiment between Japanese and nonJapanese users, and objectively analyze the difference in the information seeking behavior between the two groups by measuring their gaze and action. In addition, by focusing on the hyperlinks to the correct information, we examine the factors that inhibit the optimal search behavior of users.

In waste plants, we need to predict unstable combustion and control it appropriately for keeping combustion state stable. We have proposed a method for computing a stability degree based on fuzzy relational maps of many sensors, where we use only sensor values at the same time. In this paper, we propose a method for predicting combustion states after a certain time with multiple fuzzy relational maps of sensors. With the maps we evaluate samples from data for the maps and calculate the indices for the combustion states. We evaluate unknown data to predict combustion states after a certain time using the indices. We applied this method to data in one year of a waste plant. The result shows that multiple fuzzy relational maps can predict stable combustion states with very high accuracy and an abnormal combustion state with slightly high accuracy.

Recently, due to the widespread effects of Covid-19 infection, social communication, such as person-to-person, person-to-society, or social connection, is being diluted. As one solution to the problems, it is expected to create new social communication utilizing advanced technologies such as ICT and AI. On the other hand, the Cabinet Office is advocating a concept of ”Super City” that realize new lifestyles and businesses for fundamentally changing the state of society by AI. The authors have established a research unit of ”Health Smart Network” at Research Institute for Socionetwork Strategies (RISS), Kansai University as a research base to meet these social demands. Specifically, the purpose of this research unit is to make health promotion services smarted with ”eHealth + AI”, create new social communication, and contribute to extending healthy life expectancy and sustaining healthy life. In this presentation, we will outline the concept of the research unit of ”Health Smart Network” and introduce various researchs on the theme of smart health promotion for the progress of future research.

Internal shock such as the aging of the population and changes in lifestyle, and external shock such as the spread of infectious diseases (COVID-19) and changes in the economy, have a significant impact on health and the health system that supports it. There is a need for a system that can flexibly respond to these internal and external shock and support sustainable health. Currently, as the population ages, the need for medical and nursing care resources is increasing, and the social burden is becoming greater. With limited resources, health data is expected to be utilized to solve health issues associated with aging. Medical receipt data and long-term nursing care data held by local governments and other insurers have a large sample size, but the number of items is small, making it difficult to link them with external data. On the other hand, data acquired at the place of daily living requires a large amount of effort to acquire, but can be linked with external data to design behavioral interventions that are more in-depth into daily life. We are currently working on the “Tekuteku Beacon Project,” which aims to nudge health behavior change by acquiring behavioral data using IoT and smart tablets and linking it with medical and longterm nursing care data at facilities where independent elderly people live. In this session, we plan to introduce the behavior change interventions that can be designed based on the behavioral data of the elderly and the remaining issues.

In this study, we developed a prototype of a system that plays music for a group of people by converting their physical movements into sounds. The aim of the study is motivating them and creating a sense of unity during group fitness activities. The system uses sensors to detect the user's physical movements and converts the movement patterns into tones and pitches. Using this system, users can generate melodies from their body movements. In this study, we conducted a basic study on the effect of creating harmony by having multiple users take charge of different performance parts, and the system converting group movements into an ensemble.

The authors have been developing a design, training and evaluation application with a userfriendly operation interface for CNN (Convolutional Neural Network), CAE (Convolutional Auto Encoder) and SVM (Support Vector Machine), which can be used for the defect detection of various kinds of industrial products even without deep skills and knowledges concerning information technology. When the Grad-CAM is applied to visualizing interested areas affecting the classification results, different areas not relating to target defects are sometimes mapped majestically. In this presentation, the visualization performance of defect areas using the Grad-CAM is tried to be improved. Before learning process, all images in training data set are preprocessed by a proposed masking method, in which not-interested areas in each image are replaced with randomly generated mask patterns. The effectiveness and promise are observed through visualization tests of defect areas using the Grad-CAM.

The authors have been developing a design and training application with a user-friendly operation interface for CNN (Convolutional Neural Network), CAE (Convolutional Auto Encoder) and SVM (Support Vector Machine), which can be used for the defect detection of various types of industrial products even without deep skills and knowledges concerning information technology. The application is required to have a visualization ability of small defects which would be the causes of classification results, however, it seems to be not easy to provide such a promising function as clearly identifying the position of defect. In this presentation, CAE is applied to the visualization and position detection of such small defects included in images of industrial products. The effectiveness and promise are evaluated through visualization experiments of defect areas included in test images.

In action selection policy during deep reinforcement learning, it is possible to balance exploration and exploitation efficiently by considering the selection frequency of state action pairs. However, when the similarity of states is also learned in parallel, it is difficult to accurately count how many times each state has been reached in the past. In this paper, we propose a new method to estimate the value of each state in consideration of the balance between exploration and exploitation by constructing a network which estimates only whether or not the state has been reached in the past but has no reward. The frequency of state reached should simply increase as learning progresses, so we set such a function. The policy takes into account the mean and variance of the beta distribution constructed from reward values and their experience values. The effectiveness of the proposed method is confirmed by numerical experiments.

Ad image search system based on saliency map information is proposed for ad image designers to search close layout banners to their thinking to make. Clustering method is applied to improve its performance to search feature. Four data sets which are applied to performance test for DINet, SimpleNet and UMSI, shows that three models are applicable to ad image search system for each image types. The proposal will be used to ad rep business company to improve their work efficiency by leading new designer to senior designer with past designs and information as like a created date in a company.

Advertising movie evaluation system with UNISAL based saliency map estimation which is customized via transfer learning is proposed for movie creators in an advertising agency company to improve movie quality from viewpoint of advertising effect. The proposal provides saliency map of advertising movie candidates as objective viewpoint of audiences as like an internet user for implementation of smooth discussions with clients in video production process. Internet user’s viewpoint is shown clearly as movies with saliency map is output by the proposal, which is confirmed via two actual movie directors’ evaluation. The proposal opens new stage in advertising movie production process such as confirmation of users’ viewpoint for advertising movie blueprint and leads more effective directions.

Keyword extraction method from customer web behavior log data consists of anonymized user ID, timestamp data related with user ID, and URL where the user has accessed; is proposed to realize trend extraction with few contexts’ information in user activity logs by using naive bayes learned extra-dictionary for classification of document category and k-means to extract topic in each category. The proposal provides effective complement context information to customer web behavior log data, which is scattered in many topics, implements keyword extraction where a keyword includes word context in short text such as search queries in URL, and realizes trend information obtaining for internet advertise with low cost and short computation time. Typical word becomes mean vector for each cluster, which is a specific topic in given any period to analyze, is confirmed in experimental clustering trial applied to search queries in customer web behavior log data for seeking YouTube movies. The proposal provides more effective marketing direction via rapid observation of the trend and realizes variable and interactive digital communication for 5G era.

It was pointed out that because Japanese youth are hesitant to self-disclose, it is difficult for them to form intimate relationships. Therefore, We focused on personal nostalgia as a way to encourage deeper self-disclosure. Personal nostalgia is a kind of nostalgia based on personal events and includes narratives. In this research, We hypothesized that evoking personal nostalgia during conversation can encourage self-disclosure and improve intimacy, and conducted two subjects experiments. The results showed that intimacy increased the most in the conditions where the topics evoked personal nostalgia for both subjects. In addition, a system that recommends topics that evoke personal nostalgia for both was not related to the improvement of intimacy, but it was useful in determining the topic during conversation.

Rough-set-based interrelationship mining is a framework for mining characteristics between two different attributes using attributes called interrelated attributes. Interrelated attributes represent comparison of values of two different attributes, and adding these to data, it is capable to extract characteristics between attributes within a framework of rough-set-based data analysis. However, when there are numerous interrelated attributes we can consider, if adding all of them, then it requires a lot of computation time. Therefore, we propose a selection method of interrelated attributes based on representation ability.

In order to model preference relation from data sets, dominance-based rough set approach (DRSA) was proposed. In the setting of DRSA, classification involving a preference order is expressed by criteria, and the inconsistency in the classification is modeled by approximations of the theory of rough sets. In variable precision DRSA (VP-DRSA), a certain degree of the inconsistency is tolerated based on a given threshold. However, a method to determine an appropriate value of the threshold is not developed. The approximations of VP-DRSA can be interpreted by empirical risk minimization. In this paper, we study a threshold adjusting method based on that interpretation.

Rough-set-based interrelationship mining, proposed by the authors, aims at extracting characteristics based on comparing attribute values of different attributes at the same object. The extracted characteristics are represented as interrelated attributes. We have proposed an approach of extracting interrelated attributes that represent characteristics of multiple decision classes simultaneously. In this paper, we report various experiment results of this approach.

Evolvability of an evolutionary algorithm is considered to be an ability of the algorithm to keep producing useful phenotypic change in a dynamic environment while being robust against harmful phenotypic change. So, the evolution of evolvability is to acquire the higher ability. In this paper, we propose a genetic code for achieving the evolution of evolvability. The proposed genetic code uses an algorithm to generate graphs following a power-law which is said to be robust against random change. We show that the genetic code indeed achieves the evolution of evolvability though dynamic optimization.

Interactive evolutionary computation (IEC) is one type of evolutionary computation which includes a human user as an evaluation system. When a human as an evaluator repeats evaluation of an individual (solution candidate), he/she could be aware of his/her preference more deeply. Also, if we can reveal features of genotypes related to the awareness, it would be possible to produce better individuals including the features for him/her. In the paper we investigate if a person recognizes his/her preference well when he/she conducts an independent run for a given IEC task sequentially multiple times. We focus on mutual information of a pair of loci (positions on an individual) as the feature. The results show that most people can be aware of their preference well through sequential multiple runs and the feature corresponds to the awareness well.

Human-based evolutionary computation (human-based EC) in which humans implement evaluations and all evolutionary operators can be used for problem solving in human societies. In human-based EC, multiple solution candidates are asynchronously and in parallel produced and improved by many people on the Internet. People who join this problem solving do not have few temporal and spatial limitation. Meanwhile, a pattern language is a collection of arts represented by words for a particular human activity, and therefore, can be regarded a shareable solution to a problem in human societies. Pattern languages are synchronously and in series created and improved by many face-to-face people. People who join this problem solving have strong temporal and spatial limitation. In this paper, we design human-based EC for a production task of pattern language. Then, we discuss how the original limitation is relaxed due to the use of human-based EC and at the same time what new problems are caused by that.

In this study, we tried to understand the relationship between the thermographic images and the biological information obtained from various sensors during the exercise, and to predict the biological information from the thermographic image. The exercise adopted in this study is step exercise. Step exercise is an exercise using a stepstool like climbing and descending stairs that is performed in daily life. Specifically , the step exercises with a constant rhythm were performed, and thermographic images and biological information during this exercise were acquired as data set. The relationship between the thermographic image and biometric information was learned by LSTM using the acquired data set. After that, the heart rate, body surface temperature and the core body temperature were predicted using the learned model, and the prediction accuracy was verified.

This paper proposes a data analysis method for detecting abnormal breathing events during sleep from respiratory curves. The proposed method segments a respiration time series using Singular Spectral Transformations and classifies resulting partial respiratory curves based on their features consisting of autoregressive coefficients and respiratory amplitude ratios. Applying it to a polysomnography dataset confirmed that the proposed method could extract temporal patterns characteristic of respiratory abnormalities such as apnea and hypopnea.

Recently, we have been increasing interest in ensemble learning. In particular, Bagging and Boosting are useful methods. Bagging is a method of learning more than one classifier independently for training data, and Boosting is a method of learning more than one classifier dependent on each other. In this paper, we propose a new ensemble learning ”pdi-Boosting,” which inherits virtual data and fuzzy rules between classifiers of pdi-Bagging. In pdi-Boosting, the discrimination rate is improved, and the robustness against noise data is also improved because of increasing the amount of data due to the inheritance of virtual data. We define here the inheritance method for virtual data and fuzzy rules, and formulate pdi-Boosting. We will also discuss the usefulness of pdi-Boosting.

It is not easy to keep exercising, and it is necessary to have a system to maintain the motivation to exercise.In this paper, we propose an exercise support system focusing on the cross-modal effect in order to support users to continue exercising.We consider that using the cross-modal effect to present a simulated sense of force expected to reduce the load of exercise and to improve the motivation for exercise.Therefore, we propose a system to reduce the load of squatting exercise, and verified the effects of reducing the exercise load and increasing motivation.As a result, although the proposed system did not directly contribute to the improvement of exercise ability, but it improved the motivation to exercise.In addition, it was found that the effect of the sensory presentation by the cross-modal effect differed from subject to subject.Future research should focus on the evaluation of muscle activity and clarify the factors of individual differences.

In order to improve the quality of life of people with upper limb deficiencies, electric prosthetic hands with grasping ability are being developed.However, it is difficult to implement complex grasping motions in conventional electric prosthetic hands because the control of the grasping motion is manually programmed. Therefore, we propose a method to acquire the grasping motion of a motorized hands prosthetic hand using imitation learning. In this study, we construct a simulation environment of a five-fingered electric prosthetic hands using Unity, and aim to acquire the grasping motion by imitation learning.

A study on communication robots has become popular in recent years, and the needs of an intelligent human-machine interface are increased. Facial expression leads itself to a better understanding of human emotions because it plays a crucial part in inter-human communication. Facial expression contains various patterns, and unlearned new patterns appear over time. Therefore, a recognition model needs to adapt to new patterns and evolve. In previous studies, a method has been proposed that enables additional learning of new facial expression patterns using a threshold value called a vigilance parameter. When additional learning was performed, the average accuracy was less than 90%, and it was found that sufficient average accuracy could not be obtained. In this study, we analyzed the tendency of the training data to contribute to the improvement of recognition accuracy.

Policy gradient methods such as REINFORCE algorithm, which express the gradient function of expected rewards without using value functions, need not assume that policies of agents and environmental models of rewards and state transition probabilities have Markov property when applied to multi-agent systems. One of the policy gradient methods of this kind uses an objective function, which aims to be minimized in order to determine an action, as an energy function of Boltzmann selection corresponding to the policy. It has been shown that the objective function can be flexibly constructed. On the other hand, reinforcement learning in a multi-agent system has the state-explosion problem. As one of the effective measures to avoid the problem, a method of approximating the value function with a Boltzmann machine has been proposed. In this paper, we first propose a policy gradient method that approximates the objective function in the policy expressed by Boltzmann selection with the energy of the Boltzmann machine. Second, we propose a more efficient method that approximates the objective function with the energy of a modular structured restrictive Boltzmann machine. As a result of the experiment to a pursuit problem, it was possible to learn appropriate policies with a small number of parameters by both proposed methods. Furthermore, the second proposed method managed to significantly reduce the computational cost required for the learning compared to the first one.

In this paper, a new design method of a static output feedback control for discretetime Takagi-Sugeno fuzzy systems is proposed. This design method is based on a Lyapunov function of multiple sum type, which can relax output feedback control design conditions, and gives a wider stabilizng region than the previous method. At the end of the paper, a numerical example is given to compare our method with the other and to show the effectiveness of our proposed method.

This paper proposes a tracking controller based on input-output feedback linearization for discrete-time nonlinear systems using a piecewise multi-linear (PML) model. In this paper, we construct the PML models for the error system of the tracking control. The input-output feedback linearization is used to transform each PML model into one linear system known as the Brunovsky canonical form. Thus, the whole PML system can be stabilized by a feedback linearized controller. An Example is shown to confirm the feasibility of our proposal by computer simulation.

This paper is concerned with the problem of guaranteed cost control for Takagi-Sugeno fuzzy systems. Since all state is not necessarily obtained in a real system, we introduce the output feedback control design with guaranteed cost. A guaranteed cost control not only stabilizes systems but also considers a control performance. In this paper, output feedback control design of controller with guaranteed cost is proposed. Finally, an illustrative example is given to show the effectiveness of our propose.

It can be an effective approach for SOM learning algorithm applied to a large amount of data to adopt a parallel and distributed computing method. There are two approaches to the parallelization of SOM. One is by dividing the learning dataset and another is by dividing a SOM’s competitive layer. In this paper, the way of implementation of each two kinds of parallelized SOM and its performance evaluation are presented.

A front and back simultaneous display of the spherical SOM (Self-Organizing-Maps) was newly developed. This makes it possible to display very clearly that the phase indication on the back side of the spherical surface was unknown up to now. Significance analysis by spherical SOM was added as a further analysis method. The data on new corona measure were analyzed by the above method. The details of the data were analyzed in detail.

The professional career of female medical doctors is influenced by various life events that can lead to early retirement: marriage, childbirth, childcare, nursing, etc. In recent years, efforts have been launched to have female doctors resume their clinical practice. In order to find clues to solve the problem, we used survey data from doctors who became mothers and identified their motivation using the spherical SOM and the Residual Sum of Squares(RSS).

The traveling salesman problem has long been studied as an optimization problem that connects multiple cities with the shortest path. Angeniol published the method of SOM for the first time in 1988. Usually, a ring-shaped one-dimensional SOM is used to solve the problem. In Japan, Fujimura et al. and other researchers have studied the problem by SOM and attempted to shorten the calculation time. They developed inertial terms, eliminating intersections and learning rate functions, etc. to get shorter path and to shorten the calculation time. In this research, we applied these ideas to the traveling salesman problem while maintaining the prototype of SOM as much as possible. The subjects were att48 and att532, for which the shortest distance was already known. In the att48, the shortest path was obtained by SOM. We will show the advantages and disadvantages of the SOM method.

Recently, educational support robots have been attracting attention. In this study, we focus on a teacher-type robot that teaches learners how to solve problems. The Previous studies teacher robots do not provide learning support autonomously, but instead provide it by the learner’s button operation. This creates a situation where the learner repeatedly operate the buttons, which is likely to reduce the impression of the teacher-type robot. To prevent this problem, we think it would be effective for the robot to estimate the learner’s state of perplexion and autonomously provide learning support at the optimal timing, rather than through the learner’s button operation. By providing learning support autonomously, we think that smooth interaction can be achieved, and the learner’s impression of the robot can be improved.

Recently, educational support robots have been attracting attention. In this study, we focus on partner-type robots that learn together with learners. In the previous study, the partner-type robots perform collaborative learning by solving problems alternately with learners. However, as the learning progresses, the learner feels bored with the collaborative learning with the robot, and the impression of the robot decreases. To prevent this problem, we believe that it is effective for the robot compete with the learner for correct answer of the problems in collaborative learning. We believe that the competition of the robot make the learner feel strongly that they is learning together with the robot. Therefore, the learner does not tired of the collaborative learning with the robot and improves the impression of the robot. In this paper, we develop a partner-type robot that compete with the learner for correct answers in the collaborative learning. Moreover, we examine the impression of our robot through the experiments on college students.

This research aims at the realization of swarm flight control which is useful in the field of agriculture and amusement and so on. The swarm flight control simulation according to human operation using the flocking algorithm Boids well-known for the group behavior simulation of birds and fish was performed in this research. Boids is a kind of computer algorithm expressing swarm flight of birds based on three basic rules. In this research the parameters of three basic rules of Boids were tuned by GA. Moreover the experiment of evaluation for the operationability was performed after an operator controls one or multiple drones.

In this paper, danger level of right-turn driving at traffic light intersections is estimated by using SWLDA, Step Wise Linear Discriminant Analysis. SWLDA is one of the linear discriminant analyses that derives a regression equation, and selects the combination of explanatory variables with the highest accuracy from a number of them given. The variables selected by SWLDA are considered as the dangerous factors, we can therefore analyze the factors. We consider that quantifying the danger level when the vehicles turning right at a traffic light intersection reduces the accident rate, because drivers can be aware of the danger of his/her own driving.

In this paper, we measure the driver’s hand movement by using wrist-worn accelerometers, and the activity of the driver is estimated by means of machine learning. In order to evaluate the accuracy of driving behavior estimation, the driving simulator was used to measure driving behavior, and data for a total of 10 participants were collected. The estimated accuracy was obtained using the kNN method, and the F value was 74.5%.Furthermore, we have shown both the behavior that can be detected by the wearable sensor and its estimation accuracy.

In this research, multispectral images and RGB images are acquired by a camera mounted on a large industrial drone and analyzed by deep learning appropriately, and as a result, it contributes to the quality prediction of agricultural products. Specifically, this drone flies at a height of 5 meters or less above the ground to take an image of a tea plantation, and a method of performing deep learning by extracting feature points by wavelet expansion. This is a research fields, so called “precision agriculture”, can be said that it is highly novel. In a general technique, a multispectral camera is taken from a height of 15 meters or more above the ground and processed into NDVI images or the like to check the condition of the field. However, many agricultural products do not meet the standards. Tea is not the fruit but the leaves themselves is a crop, and it can be said that each tea leaf is small and has a similar color, so the results of image analysis cannot be expected. We think that this can be solved by devising the imaging method, the type of image signal, and the analysis method.The case is more difficult than other crops and is a widely applicable study.

Many studies for privacy preserving data mining have been done. Secure Multiparty Computation (SMC) has been introduced for privacy preserving data. SMC was applied to some statistical computation methods. In previous papers, we proposed synchronous Back Propagation (BP) learning for SMC. In this paper, we propose the more general BP learning method.

This study aims to improve the performance of unsupervised morphological analysis using NPYLM for minority languages. Conventional methods require a large amount of data for training, but the amount of data for minority languages is limited. So far, we have tried to improve the performance of unsupervised morphological analysis by using the ”replacement” method, in which words that are correctly analyzed are replaced with different symbol types even if the amount of data is limited. As an improvement of the ”replacement” method, we have also studied a ”limited replacement” method based on TF-IDF under the assumption that words should be replaced and words that should not be replaced. In this paper, we aim to improve TF-IDF’s performance by making further improvements in its operation. As a result, the F-value of TF-IDF is greatly improved, enabling us to extract word candidates efficiently even from documents consisting of unknown words.

In recent years, there has been a growing interest in self-actualization. In order to achieve self-actualization, it is necessary to be highly motivated in learning. In this paper, we propose an environment in which learners learn with multiple agents with different characteristics. We propose a learning environment in which learners learn with multiple agents with different characteristics. The learners can take turns giving quizzes to the agents, and the agents can give advice to the learners while observing them solve the quizzes. By using the agents, the learner can perform cooperative learning alone at any time. In addition, by adding a game element to the learning method, we gave the learner an incentive to learn and a goal to cooperate with the agent. In the experiment, we evaluated the proposed system to see if it was useful for learning. As a result, we found that most of the subjects felt that the proposed system facilitated their learning.

In recent years, the number of people who use bicycles in Japan has been increasing rapidly, andmanytrafficaccidentshavebeenoccurring. Therewere67,000trafficaccidentsinvolvingbicyclesin 2020, accounting for 21.9% of all traffic accidents in Japan, and this ratio is increasing every year. In this paper, we propose a bicycle collision prediction system using object recognition.

In order to improve the productivity of agricultural crops, research is currently being conducted on the modeling of growth conditions. The leaf area is important for modeling tomatoes. However, the current measurement method involves the destruction of leaves, and it is not possible to relate the destroyed leaves to future growth of tomatoes. In this study, we propose a system that non-destructively estimates leaf area from depth cameras using deep learning.

Tamagawa river in Akita, Japan, has possibility to impact surrounding environments because an inflow of acidic water including hydrous ferric oxide, arsenic compound, and others. Our previous studies were working for water quality analysis using satellite remote sensing techniques to Lake Hosenko has water source from Tamagawa river. Concretely, fuzzy c-means clustering was used for creating estimation maps of water quality. As the results, the FCM was useful for understanding the pollution due to hydrous ferric oxide in the lake. However, the previous studies used only satellite data for creation of estimation map. Therefore, the purpose of this study is to discuss how to specifically obtain state information of water quality and to develop a method for creating estimation map of water quality using neural network with satellite data and topographical information of Lake Hosenko such as elevation as input-features. In this paper, water quality maps were created using each band of satellite data and elevation of Lake Hosenko as input-features.