Along with the development of super-aged society, cases are increasing that extend healthy life expectancy for the elderly. Dementia is one of the factors that hinders the extension of healthy life expectancy. Therefore, it is important to prevent dementia and maintain and to improve cognitive function. The number of cases that adopting electronic sports (esports) is increasing because it is effective in maintaining and improving the cognitive function of the elderly. Quantitative detection of stimuli and emotions obtained from esports can contribute to the prevention of dementia in the elderly. In this paper, we examined the relationship between information of complexion and emotions and interests during esports for the elderly.

The purpose of this study is to implement a human consciousness model based on Jung psychology as a learning model of a machine learning agent so that the agent can acquire individual character from the result of its own behavior. While existing AI agents have been given characteristics by incorporating human-like behaviors and reactions according to the environment, this method make it possible to realize a variety of AI agents without human labor. The consciousness model has a part that corresponds to the instinct including the three major human desires, and this is one of the behavioral factors of the agent. In this paper, we focus only on "appetite" and use reinforcement learning to learn behaviors that satisfy appetite. The parameters, learning environment, and data exchanged to satisfy appetite are implemented based on Jung's typology. Multiple agents are prepared by changing the initial parameters, and learning is performed for each. Each trained agent is activated in the production environment. The behavior and accumulated data of each agent are compared and evaluated with the results of manual operation by humans. As a result, we verify whether we can obtain a difference that can be said to be the individual character of the agent.

This study forecast Japanese confirmed cases of the novel coronavirus in order to assist in decisions. There are statistical models and machine learning models for forecasting the time series data. Statistical models performed better than machine learning models, which was confirmed by the experiment results of the comparing. Therefore, this study forecast confirmed cases of the novel coronavirus with SARIMA(Seasonal AutoRegressive Integrated Moving Average) and RNN( Recurrent neural network compares each model with RMSE (Root Mean Square Error). As the result, RNN(with vector inputs) is better than machine learning models.

There has been a lot of research in the medical field aimed at substituting artificial intelligence technology for diagnosis and analysis by medical professionals. Among them, we focused on the analysis of immune cells. Immune cells have been analyzed manually, and cell selection in particular has been a heavy burden for workers. Although Recognition Frequency Space has been proposed to automate the selection process, it is difficult to obtain the neighboring cells individually. In this paper, we propose a selection method to improve the selectivity of neighboring cells using Grad-CAM and Grad-CAM++. As a result of comparison with the previous methods, the selection results using Grad-CAM showed that each of the three neighboring cells could be selected individually. This is because of the higher performance of Grad-CAM in performing precisely weighting compared to previous methods.

In recent years, wild macaques have been damaging crops in mountainous areas in Japan. Since monkeys live in groups, it is necessary to capture all the monkeys in a group to prevent damage. Also, once a monkey is released from a trap, it learns and the group of monkeys will not approach the trap. Therefore, we need a system that can detect when a group of monkeys enters a trap by using image recognition. Also, unlike other animals, monkeys are active during the daytime in search of food. Therefore, the image recognition system needs to be designed to take into account the effects of sunlight outdoors. The goal of this study is to detect a group of monkeys. In this paper, we investigate a method of detecting and measuring individual monkeys as a way to detect a group of monkeys from an image.

In Kochi Prefecture, Yuzu is widely cultivated, and its cultivation managements using information technology have been getting attention in the agricultural field. In order to predict shipping by automatically counting crops, location detection of fruits using CNN has been proposed. Here, there is a problem that the cost of creating supervised data for supervised learning is high. This paper proposes improvement methods using Noisy Student based semi-supervised learning with a teacher model and a student model as a way to utilize easily collected unsupervised data. We examine three improvement methods: scale expansion, pseudo-label generation using test time augmentation, and transfer learning to the student model using the weights of the teacher model. As prediction models, we use two trained models of YOLO v5. The transfer learning using only the generated teacher data results in mAP of 38.1%, while the transfer learning using scale expansion results in 70.6%.

Onomatopoeia can easily describe sounds or state of things. Onomatopoeia is believed to follow sound symbolism. Sound symbolism is a property that particular sound can make people imagine particular impression. Urata et al. focused on sound symbolism and proposed an onomatopoeic thesaurus map to visualize semantic similarity among onomatopoeias. The map is composed of outputs of a middle layer of an autoencoder that has learned sound symbolism. Urata et al. hired small number of limited onomatopoeias to construct the map. Besides that the limited onomatopoeias took not only a XYXY type but various styles. As a result, there is a possibility that the map are only able to visualize limited semantic similarity among onomatopoeias. In this paper, we prepared many XYXY-typed onomatopoeias to construct the autoencoder. Using the autoencoder, the experiment attempted to verify the map regarding visualization of sound symbolism. We formulated a hypothesis regarding sound symbolism based on Japanese linguistic knowledge. Most of the results supported the hypothesis, however some showed a different trend. We discussed on that.

In educational support systems using robots, gaze information is considered to play an im- portant role. There have been many reports on eye gaze for children with regular development or with developmental disorder in learning support systems using robots. On the other hand, there are few find- ings for children with “mild developmental disorder”. Tabot Egg is an educational robot that aims to support the education of children with mild developmental disorder. Until now, Tabot Egg has not been equipped to collect and analyze gaze information. Therefore, we have implemented a function in Tabot Egg to analyze eye gaze data. In this study, we installed an image sensor on Tabot Egg and collected gaze information from test subjects to predict gazing regions. The experiment in this paper set up two classes of the gazing regions: a robot region and a non-robot region. The classes were predicted by inputting the gaze information obtained from the image sensor. As an experimental result, the mean accuracy achieves 75%. Based on the result here, we also investigated the possibility of a multi-class classification problem, i.e., high-resolutional gazing-region prediction.

In this study, we propose a robotic model that makes decisions based on group norms as a group member that have interests. In a field of social robotics, it is important that robots naturally participate in human communities and behave in a human-like way. In our previous studies, robots behaved socially in human-robot groups by obeying group norms. However, our previous studies did not consider interests of the group members. In general, a beneficial behavior for a member might put other members at a disadvantage. Due to conflicts of interests in a group, each member tends to behave in a way that requires others to be fair while trying to gain more benefit. Therefore, under group norms to behave with fairness in mind, each group member pursues their interests. In human-robot groups, the robots also need to adapt their behaviors to the group, considering the implicit norms of fairness. In this study, we investigate how an agent using the proposed interest-aware model behaves in interactions with humans. In the proposed experimental scenario, we confirmed that each group member’s idea of fair behavior under the interest relation appeared as a group norm.

In this paper, we propose a decision-making method for robots based on robot ethics. Specifically, we applied the two-level theory of utilitarianism, which consists of SYSTEM 1 (intuitive level) for quick decisions and SYSTEM 2 (critical level) for slow but careful decisions. In this paper, SYSTEM 1 is a set of heuristically determined responses, and SYSTEM 2 is a rule-based discriminator. An impression evaluation experiment was conducted, and the effectiveness of the proposed method was suggested.

The purpose of this study is to estimate the state in the human brain when image stimuli are given, and in particular, we will verify the estimation accuracy and visualization results by various deep learning models for image processing as a working model. The human brain activity when images are shown to a subject is observed using fMRI, the same image is input to various deep learning models for image identification, and the representation of those intermediate layers is regressed to the brain activity state. By this, we examine the difference in estimation accuracy for each deep learning model through visualization, and investigate the characteristics of deep learning models when estimating the human brain activity evoked by visual stimuli.

It is known that not all parts of the visual patterns are not equally perceived, but convex parts are perceived more saliently than concave parts (Subirana-Vilanova & Richards, 1996, etc.). However, it is not a trivial to determine where are the salient parts, because convex / concave parts become clear through the analysis of the whole pattern. This study proposes a simple model to extract perceptually salient parts on the shape cognition based on the model of figure-ground separation proposed previously by the author. The function of the model was confirmed by computer simulation.

There have been many attempts to recognize emotions based on machine learning and biometric data, such as EEG. In these attempts used emotion model has a significant impact on the estimation of the emotions. In this study, we examined the validity of the models by comparing the pleasantness estimation based on EEG and the internal emotion estimated by the three representative psychological emotion models; Russell’s emotional dimension model, Plutchik’s Emotion Wheel, and Ekman’s basic emotions. The results showed that the pleasantness estimated by Plutchik’s model was significantly different from the pleasantness estimated by other models and EEG, and that the pleasantness estimated by Ekman’s model was closest to the pleasantness estimated by EEG. The results suggest that the model including no complex emotions such as Ekman’s model most closely matches to the objective index of brain activity in the viewpoint of the degree of pleasantness and displeasure.

Meditation has been used to control the mind since ancient times. Recently, mindfulness-based meditation research is a typical meditation research. It has been scientifically shown to reduce stress and anxiety, and it is used for the purpose of improving concentration or mind condition in psychiatry place and even in corporate training. Previous studies reported that the appearance of α-waves in the frontal region and the appearance of θ-waves as features of EEG signal during meditation. On the other hand, the appearance of θ-waves in the midline of the frontal region has been reported as a characteristic during concentration, suggesting a relationship between meditation and concentration, but there is no report directly clarified the relationship. In this study, we identified the “ meditative state ” by an objective physiological index and verified the increase in concentration due to the meditation. As a result, the intensity of θ-wave power during meditation and the intensity of θ-wave power during task performance were correlated, directly suggesting the enhancement of concentration by meditation. It was suggested that the parasympathetic system became dominant after meditation, and it was confirmed that the meditation state could be defined by the EEG and the electrocardiogram. Thus, the meditative states were confirmed as EEG and heartbeat states, we are currently investing whether these states can be reproduced by classical conditioning.

It is difficult to find an appropriate evacuation route in complicated buildings and underground malls. Therefore, evacuation guidance systems using sound and light has been proposed. The method of the emitting sound was proposed as one of them. In this evacuation guiding method, several influential factors such as speaker setting configurations and sound emitting methods affect the recognition of the emitting direction of sound. In this paper, we discuss the sound emitting configuration to improve the performance of the sound direction recognition through the experiments changing the emitting configurations of audio speaker distance and emitting time interval.

Thanks to a continual learning capability and superior classification performance, Adaptive Resonance Theory (ART)-based clustering has been actively studied. Especially, ART-based topological clustering algorithms have shown superior clustering performance than other clustering algorithms. However, those algorithms have a data dependent parameter, called a vigilance parameter, which has a significant impact on the clustering performance. This paper introduces an automatic vigilance parameter estimation method to an ART-based topological clustering algorithm. Experimental results show that the proposed algorithm achieves superior clustering performance on a 2D synthetic dataset and real-world datasets compared to conventional algorithms.

With the advancement of artificial intelligence technologies in recent years, research on intelligent robots has progressed. Communication robots that engage psychologically with people, in particular, are gaining traction. Hand gesture recognition is important in human-robot interaction because it allows us to comprehend diverse human gestures and their intention. Traditional deep learning approaches need to keep all prior class samples in the system and require model training from the start by combining prior and new examples, which consumes massive amounts of memory and dramatically increases computation cost. In this study, we proposed a method called Incremental Recurrent Kernel Machines (IRKM) that mimics the human lifelong learning process that can continuously learn new gestures without forgetting previously learned gestures. The proposed method consists of two hierarchical memory layers: i) Episodic Memory and ii) Semantic Memory layer. The Episodic Memory layer incrementally clusters incoming sensory data as nodes and learns fine-grained spatiotemporal relationships of them. The Semantic Memory layer adjusts the level of architectural flexibility and generates a topological semantic map with more compact episodic representations based on task-relevant inputs. The generated topological semantic map reflects the memory of the robot in which it is utilized for gesture recognition.

In the field of clustering, which is a technique for automatically classifying and summarizing data, rough clustering, which deals with the uncertainty of belonging to clusters by introducing a viewpoint of rough set theory, has attracted attention. Collaborative filtering based on rough clustering such as rough C-means (RCM) has been proposed and reported to be effective. However, the size of data used in a collaborative filtering task is generally very large, and it is considered to be difficult to apply conventional algorithms to the task. In this study, we propose online RCM introducing online learning and experiment with its application to collaborative filtering. Furthermore, we verify the effectiveness of the proposed method through numerical experiments using real-world datasets.

Growing Neural Gas (GNG) based space perception systems have been proposed by many researches since the GNG can learn a geometric structure and generate a topological structure simultaneously. However, these proposed methods assumed the simple perceptual task and cannot apply the multi perception tasks because theses proposed methods use the background extraction. For solving this problem, our previous research proposed Region of Interest GNG (ROI-GNG). ROI-GNG can build concentrated/distributed geometric structure according the degree of attention by controlling the discount rate of the accumulated error according to the degree of attention. However, ROI-GNG has a problem about the stability of the node addition and deletion capabilities. In this paper, therefore, we propose a modified ROI-GNG based on GNG with targeting (GNG-T) that can stably learn the geometric space of the vision sensor by utilizing the d-shortest confidence interval to the algorithm related with the node addition and deletion for improving the stability. In addition, we designed the parameter of the related importance for applying modified ROI-GNG to image data. Finally, we show experimental results of the proposed method and discuss the effectiveness of the proposed method.

This research proposes a Growing Neural Gas based topological environmental map building method from a metric map with high resolution map for using the self-localization. Next, the path planning method is proposed by utilizing the occupancy information of the topological map. Finally, we conduct on several experiments for evaluating our proposed method.

Accurately acquiring information of product color is very important in quality control produces. Like in color scheme of industrial products and chemical inspection using colors, a method of visually comparing and measuring between object and standard color chart is generally used. However, it has some problems that need cost and labors. In online shopping, which has been increasing in recent years, the colors of purchased products is often different than those expected due to the difference of light-source color. For the purpose of eliminating the gap felt by consumers, a pixel value correction method that reduces the influence of ambient light and analyzes the color characteristics of the object is suggested for image acquired using smartphone and tablet PC. In this paper, we conducted a study of color feature analysis and material estimation using image features for three types of cloths made by different materials.

Japan has the highest proportion of elderly people in the world, and the incidence of dementia is also increasing. As a physical exercise, cognitive training, and social participation are being promoted to prevent dementia. Because e-sports can be enjoyed while interacting with many people, and it is possible to obtain greater emotional changes than the emotions evoked in daily life, it is expected to be an initiative to prevent dementia and extend healthy life expectancy in the elderly. In addition, if it is possible to obtain fluctuations in the amount of heat at the facial skin temperature by implementing e-sports, it is expected that the brain will be activated due to changes in blood flow. Therefore, with the aim of scientifically clarifying the effects of e-sports in the elderly, we examined the amount of change in skin temperature before and after doing e-sports in facial thermal infrared images.

Electronic substrates contain metals of high economic values such as gold, silver, copper, and palladium. In non-ferrous metal smelters, these metals are recycled and valuable ones are being recovered. In order to recycle the spent electronic substrates, it is necessary to estimate the valuable metal content in the spent electronic substrates in advance, and select an appropriate treatment method. In this paper, we extracted parts containing copper on a spent electronic substrate using image recognition technology, and conducted a basic study of estimating copper content using features.

Currently, systems for detecting social disruptions, including natural disasters, have problems such as being able to detect only specific disasters or small social disruptions. In contrast, by analyzing human flow data, which is data on the movement and flow of people, we can expect to build a system that can detect general-purpose social disruptions. In this paper, by using precipitation data in addition to human flow data, we analyze the effect of precipitation on human flow, and report the results of data analysis conducted using both human flow and rainfall.

Currently, the number of visually impaired people is on the rise. There are many accidents involving the visually impaired while walking outdoors, and demand for walking support systems is high. In recent years, the number of stations with platform doors on the platform has been increasing, but there are still many stations that do not have such doors, which may cause accidents involving falls. In addition, visually impaired people cannot tell whether the steps in front of them are stairs or dangerous steps such as platforms. In this paper, we propose a system to detect upward and downward steps by combining depth images and CNN models. In particular, for downstairs steps, we distinguish between progressible steps such as stairs and dangerous steps such as platforms. In this research, we take depth images using a stereo camera that can obtain depth images, and create a data set. Using that dataset and CNNs, we will build a model that classifies four classes: ” down stairs”, ”down step”, ”up step”, and ”plane”.

In this study, English sentences of “English” for the Common Test for University Admissions are examined in terms of metrical linguistics, compared to those for the National Center Test for University Admissions, and English textbooks for Japanese junior high and high school students. In short, frequency characteristics of character- and word-appearance are investigated using a program written in C++. These characteristics are approximated by an exponential function. Furthermore, the percentage of Japanese junior high school required vocabulary and American basic vocabulary is calculated to obtain the difficulty-level of each material.

In university lectures, students' post-lecture feedback varies widely, but it can be categorized. Through categorization, the first author examined what kind of feedback students give and what kind of scores they get on tests in relation to individual test questions.

The purpose of this study is to develop a system for detect topic transitions in a chat dialogue in order to grasp the context, which enabling to reduce conversation breakdown. Words used in utterances are vectorized using Word2Vec, and the similarity between utterances is calculated by cosine similarity. Then, suppose that the topic transitions when the similarity between utterances becomes low. Using an actual dialogue corpus, we conducted a subject experiment comparing the results of the system extracting topic transitions with the results actually judged by subjects, and found that the topic transitions were captured at a position close to the human judgment. However, the system presumed that there was no topic transition when similar words were used even for different topics. In addition, the accuracy was improved by adding the algorithm to use an intersection of words between distant sentences.

A blind physics teacher hopes to edit graphics independently, which are distributed to his students in his physics lecture. The physics teacher edits graphics using TikZ, which is a drawing package of LaTeX. However, it is extremely difficult for visually impaired people to think about editing process and accurately specify the positions of basic shapes such as lines and circles. In addition, since the figure is presented on the display, visually impaired people cannot check if the figure was produced correctly without an assistance of a sighted person. Therefore, we prototyped a novel object-oriented graphic description language that is easy to edit even for the visually impaired, and developed a graphic production system using the language. This paper describes the system we have developed.

In recent years, the number of patients is a tendency has been increasing in Japan with the gain in aged people. ALS patients are unable to move their bodies freely by themselves. So there are gaze- based living support systems have been marketed for some time. However, most conventional systems have a user interface that includes screen transitions, which aged people, who are prone to ALS, tend to have difficulty with. In this study, we focused on dialogue systems, which have been developed and spread in recent years. We believe that an interactive user interface using the Japanese language will eliminate screen transitions and allow even aged people to operate the system smoothly. In this paper, to develop the proposed system, we first developed a virtual home appliance operation simulation system for ALS patients. In addition, we conducted an evaluation experiment of the proposed system by having the participants of the experiment (healthy people) operate this simulation system using a screen keyboard input by gaze and a task-oriented dialogue system.

In recent years, research and development on automated driving has been conducted at the national level in countries around the world. On the other hand, according to a survey conducted by the Dai-ichi Life Research Institute, about half of the participants were concerned about automated driving, and of those, about 70% were concerned about the reliability of automated driving. In addition, previous studies have suggested that voice notifications, instead of sound notifications, would reduce the annoyance felt by drivers when switching from automatic driving to manual driving. However, there has been no detailed discussion on the design of voice notification at the time of phase release. The purpose of this study is to improve the reliability and reduce the annoyance of automatic driving. We have proposed a driver assistance agent that speaks with consideration for linguistic considerations based on the strategy of politeness theory, which is a linguistic strategy for smoothing human relations, and evaluated the acceptability of the utterance.

International standards and organizations are working to provide access to printed and electronic books for the visually impaired. As part of this effort, there is software that reduces the burden of Braille translation by using text OCR and formula OCR. However, there are very few such efforts for image data, and the figures in the literature are translated into Braille by sighted people. In this case, there is a vector format that is accessible to the visually impaired. Due to its characteristics, vector images can be easily redrawn and re-edited, and they can be read by point chart readers and 3D printers, so the use of vector images is recommended. On the other hand, not only printed books but also most electronic books are stored in raster format, and it costs a lot of money for editors to convert raster images to vector images. In this paper, we propose a method to analyze raster images and convert them into vector images using deep learning.

In the previous study, some of the present authors had analyzed spatiotemporal human brain activities during the tetrapod image naming process, and pointed out that activation of the right angular gyrus depended on its shape. There are at least two processes in naming: the image recognition process and the name recalling process. It is not known whether the activity in right angular gyrus is involved in the recognition process. The present authors analyzed brain activities in image evoked recognition of four animals using equivalent current dipole source localization (ECDL) method.

The parahippocampal gyrus is related to the immediate memory matching and the recent memory matching in higher-order recognition process in the human brain, and the right parahippocampal gyrus is related to the image recognition. In addition, some recent studies suggest that the parahippocampal gyrus performs memory matching process during the primal visual recognition process. In the previous study, some of the present authors measured electroencephalograms (EEGs) on subjects observing and recalling four tetrapod images. Then we had localized higher-order image recalling process in human brain. In this study, the present authors used single trial EEGs according to images, those are presented at first among the same images, and localized brain activities using equivalent current dipole source localization (ECDL) method. ECDs were localized to the right parahippocampal gyrus during primal visual recognition process.

In this paper, we represent speech stimuli when watching a movie as BERT embedding vectors; build a regression model for estimating the brain activity state under the stimuli using the vectors as input; and evaluate the model by means of the correlation coefficient between the estimated state and the real state. Through this model, it was confirmed that the state of brain activity when brain received stimuli of the meaning of words can be estimated.

In recent years, numerous attempts have been made to explore the decoding of brain activity. Specifically, a great deal of work has been done to capture the general correspondence between language and brain activity. In this paper, we investigate the mapping between text and brain activity data for brain decoding using BERT [1], a generic language model proposed in the field of Natural Language Processing (NLP). At the same time, we investigated the effect of brain activity on the NLP task using Brain BERT, a novel model promoted by our team. We also compared and validated different models to obtain the best bias and weights for Autoencoder and Brain BERT model to better extract the brain features.

Non-negative matrix factorization (NMF) is a basic method for analyzing the intrinsic structure of such non-negative matrices as environmental observation data, but cannot work well when datasets include some subsets drawn from different generative schemes. This paper proposes a novel switching NMF algorithm, which simultaneously estimates multiple NMF models supported by a fuzzy clustering concept. The NMF least square measure is modified by introducing fuzzy memberships of each object, and object fuzzy partition estimation and cluster-wise local NMF modeling are iteratively performed based on the iterative optimization principle.

Adaptive Network-based Fuzzy Inference System (ANFIS) is a promising model of explainable neural networks but rejection of illegal noise effects is an important issue in real application. In this research, a novel approach for introducing the noise clustering concept into fuzzy c-means-based ANFIS is proposed for robust modeling, which simultaneously considers two types of noise generation schemes of input-level noise and output-level noise.

Privacy preserving data clustering is a useful method for extracting intrinsic cluster structures from distributed databases keeping personal privacy. In this research, a novel model of extracting plane-like fuzzy clusters is proposed, where privacy preserving scheme of k-means-type model is enhanced with fuzzy c-varieties utilizing cryptographic calculation. The element-wise clustering criterion enables to derive local principal component vectors in each data sources.

In the field of fuzzy data analysis, a method of expressing the uncertainty of human thinking by using fuzzy interval data and applying principal component analysis has been proposed. In addition, from the standpoint of statistical learning theory, principal component analysis for interval-valued data has been proposed in the field of symbolic data analysis. Fuzzy interval data has been applied to these methods and their properties investigated. However, the data to be investigated is a data structure under specific restrictions, so various data structures have not been investigated. In addition, performance surveys using evaluation values concerning explainable power of the data have not been conducted. Therefore, in this paper, we purpose a method to investigate the applicability of fuzzy interval data to the interval-valued data principal component analysis proposed in the symbolic data analysis by using reproducibility, visualization, and evaluation values of the principal component analysis.

We apply an artificial bee colony (ABC) algorithm to interactive evolutionary computation (IEC) method for multimodal retrieving of candidate solutions. Previous works have proposed IEC systems using a parallel interactive tabu search algorithm (PITS) that generates multiple tabu search algorithm (TS) retrievals and hybrid a genetic algorithm (GA) that global retrieving method and a TS that local retrieving method for multimodal retrieving. However, the PITS cannot efficiently retrieve candidate solutions and has complicated algorithm. Also, the hybrid GA–TS hard to retrieve candidate solutions if the user has a more multimodal preference. Then, we propose an IEC method with ABC algorithm for multimodal and simultaneously retrieving of candidate solutions. In this study, we perform a numerical simulation with a pseudo user that imitates multimodal preferences as target individuals instead of a real user. The results show that the proposed method can retrieve multimodal candidate solutions in conditions with limited number of candidate solutions and bees.

In this study, we verified fuzzy inference method in the character coordination system using the Kansei retrieval agent (KaRA) model. In this verification, we conducted a comparison experiment between two methods: Min–max method, which is a commonly known inference method for fuzzy inference, and other method that uses the average value instead of the minimum value for the antecedent part membership value. The character coordination system learns user preferences by optimizing parameters of fuzzy inference in the system. Previous studies have demonstrated that the character coordination system is effective in terms of acquiring preference rules that are the evaluation criteria of users. However, in previous studies, it is not possible to extract rules that the user does not consider (Don’t care) in the evaluation of coordination. Therefore, in this study, we verify the effect of preference rules extraction when using Don’t care label which indicates that a feature not considered in the evaluation of coordination in character coordination system. As a result, we found that both methods can generate rules that fit users more than about 70%.

The polygon packing problem is a combinatorial optimization problem that requires placing a given set of polygons within a rectangular strip with a fixed height and an arbitrary length. And, we aim to minimize the width. However, each polygon must not overlap. In this paper, we propose a new method for converting into the rectangular packing problem using a hierarchical genetic algorithm, which combines multiple pieces and approximates them into a rectangle. Generating a rectangle is divided into steps to optimize 1) the pieces and angles to be combined, and 2) the margins when the rectangle is viewed as a rectangle. In this way, we only consider rotations of 0 and 90 degrees, which is expected to reduce the computational cost significantly. We conduct experiments using a data set with and without rotation, and compare the final result with existing studies.

Multi-objective Fuzzy Genetics-Based Machine Learning (MoFGBML) optimizes fuzzy classifiers considering two objectives: maximizing accuracy and minimizing complexity. Generally, a single shape of antecedent conditions is used for fuzzy if-then rules in MoFGBML. However, an appropriate shape of membership functions probably depends on datasets. It is also possible that an appropriate shape is different among attributes even in the same dataset. In this paper, we define six antecedent conditions, combinations of three shapes (i.e., a triangular fuzzy set, a Gaussian fuzzy set, and an interval set) and two types of partitions (i.e., homogeneous and inhomogeneous) and propose MoFGBML simultaneously using these six kinds of antecedent conditions. Through various computational experiments, we discuss the effects of using multiple shapes of antecedent conditions on the accuracy and complexity of the obtained classifiers and analyze the frequency of selected conditions for each attribute.

Due to the spread of emails and SNS(Social Networking Service), we have more opportunities to communicate with each other by writing. However, it may be hard to understand written sentences and expressions, and even their connotation. In this study, we make the facial expression of emoticons for various situations and written sentences by adopting affirmative/negative impression degree of their phrases.

Research on dialogue agents that reproduce counseling dialogues by psychotherapy is in progress. However, dialogue agents currently do not have the same psychotherapy technology as same as psychotherapists. Because of that, the dialogue control of dialogue agents depends on the prepared scenario. We aim to realize a counseling dialogue system that can promote flexible dialogue for problem-solving by introducing the dialogue stage prediction to the dialogue system. Therefore, as the first step, we designed the features suitable for predicting the progress stage of the dialogue. As a result of conducting a dialogue stage division experiment by non-hierarchical clustering using those features designed for dialogue stage prediction, the feature considering the context information and polarity during the dialogue showed the best performance with Purity 0.69, Inverse Purity 0.72, and F value 0.70. Therefore, the effectiveness of our proposed feature design is confirmed.

Local governments publish information about parliamentary activities and financial conditions on their Web sites, many of which are available in PDF document format. PDF documents have a high level of visual expression such as layout. However, it may be difficult to extract text data such as letters and word order from PDF documents because of high flexibility of their internal data representation. Since local governments publish PDF documents with their own format, the methods and difficulty of extracting text data may vary. To be useful as a language resource or Linked Open Data, it is desirable to structure documents on a plain text basis, such as JSON. In this study, we analyze PDF documents of assembly materials published local governments in Fukuoka Prefecture from the viewpoint of the method and difficulty of extracting plain text.

This research is concerned with improving the performance of unsupervised morphological analysis, which segmentsgivencharacterstrings ofanunstructuredlanguage into word units.The conventional method, NPYLM, shows poor segmentation performance for less data. In this paper, we discuss the cause of the poor performance especially focusing on low frequency words. We investigated the relationship between the percentage of low-frequency words and the segmentation performance by varying the amount of data. As a result, we conclude that low-frequency words increased impact on the segmentation performance degradation with decreasing the amount of data.

In this paper, we propose a model for predicting the lateral passing distance between pedestrians and vehicles by using the inclusion-exclusion integral regression model. In the experiment, we compare the inclusion-exclusion integral regression model with neural network and multiple regression analysis models.

Human-like conversation with gestures and verbal cues makes a contribution to provide more natural communication. In this study, we proposed an approach for robot's imitative learning in human-robot interaction. This paper presents a method of robot motion generation based on a steady-state genetic algorithm (SSGA). SSGA is one of evolutionary optimization methods using selection, mutation, and crossover operators. Moreover, we discuss the applicability of the proposed approach to incremental learning of depictive gestures in collaborative human-robot interaction. We show some experimental examples to discuss it in this paper.