Recently, evolutionary multiobjective multitasking (EMOMT) that solves multiple multiobjective optimization problems (i.e., multiple tasks) in parallel using evolutionary computation has been actively studied. In evolutionary multitasking, each task has a population to be optimized by evolutionary computation. The main feature of EMOMT is that offspring individuals are generated by not only intra-task crossover but also inter-task crossover. Previous studies show that adding offspring individuals generated by inter-task crossover to each population improves the search ability of EMOMT. In inter-task crossover, the position of parent individuals in the decision space affects the generated offspring individuals and the next population. For example, when the position of parent individuals is extremely far away in the decision space, the generation area of offspring individuals becomes very large. However, selection of appropriate parents in inter-task crossover is not well-studied. In this paper, we focus on the similarity of individuals between two populations (i.e., two tasks) in the decision space and examine the effects of different parent selection schemes on the search performance of EMOMT.
We strive to develop a dialog system which have a sense of humor in order to improve user satisfaction. We have constructed a large Japanese pun database. According to the analysis of the puns, onomatopoeia words assume an important role. In this paper, we report the analysis of onomatopoeia words in Japanese puns.
The recent rapid development of information technology has enabled us to collect data continually. The accumulated data is regarded as an important economic resource. Many researchers have studied techniques for extracting knowledge in an interpretable format from continually increasing data in view of quality and quantity. Fuzzy genetics-based machine learning (GBML) is one of the most effective methods to design a classifier to extract interpretable knowledge. By applying fuzzy GBML, we can obtain a fuzzy classifier composed of linguistically interpretable rules. However, fuzzy GBML cannot learn in a situation where training data with unknown class labels increases continually (i.e., class incremental situation) because its learning algorithm is batch learning. In this paper, we propose class incremental fuzzy GBML (CI-FGBML), which can learn in the class incremental situation. Specifically, when patterns in an unknown class obtains, CI-FGBML performs two operations: i) regeneration of rules for classifying the unknown class, and ii) reduction of training data belonging to trained classes. Experimental results show that the proposed method can learn efficiently in the class incremental situation from the viewpoint of computational cost.
Many kinds of onomatopoeia in Japanese have multiple meanings: “gorogoro” is used in a different sense, such as “thunder rumbling” and “chilling out at home.” From previous researches, it is known that determining the meaning of onomatopoeia is depend on the dependent verbs. However, there are also onomatopoeia without verbs. Additionally, the meaning determination may require extensive contextual information. In this paper, we examine the possibility of senses classification considering the appearance context of onomatopoeia using BERT pre-trained model, which is a general-purpose language model. The evaluation results of sense classification show that accuracies of “gorogoro” and “batabata” are 73.9% and 57.8% respectively. Although the classification performance was significantly different by onomatopoeia’s senses, the performance of senses which was included more in the training data was good.
The purpose of this study is to change the impression given to the user by changing the words of the sentence output by the dialogue system to intimate words or words with respect using the politeness theory known as the consideration for constructing the smooth human relation by the words and deeds of each other. Concretely, based on the politeness theory, from the chatting data between human beings, the words which shorten the psychological distance by casual talk and communication and the words which maintain the psychological distance by honorific language were extracted. Using the dialogue data, we learned the appropriate reply to the other party’s utterance using the sequence to sequence model suitable for learning the paired data such as utterance and reply. After learning, we modified the output so that the Seq2Seq model uses two kinds of words preferentially. On the dialogue system which changed learning and output, the evaluation experiment was carried out whether impression and appropriate response were carried out.
The author defined type-2 fuzzy contingency table concerned two variable, and proposed fuzzy portfolio analysis method that applied it. Also, we have reported questionnaire analysis on classes as application of it. In this paper, we propose a method to compare the order relation of 2D fuzzy vectors by 1D reduction. Furthermore, we show its effectiveness through ordering improvement items in fuzzy portfolio analysis.
This paper describes a real-time situation evaluation system for RoboCup Soccer Simulation 2D League and its application to a spectating system. The aim of developing the spectating system is to make the spectator’s experience of watching games more entertaining by expressing the excitement of the audience based on the game situation, and making it easy to understand the game situation. For this purpose, the evaluation value of the game situation called SituationScore is estimated in real time during the ongoing game, and the excitement of the game is expressed by the audio and visual effects according to the estimated SituationScore. How SituationScore is obtained using a deep learning method and also how a soccer monitor application is extended with audio and visual effects are explained.
In this paper, we propose estimation methods of the N-value, which is an index for measuring the ground strength. The proposed methods use multi-layered neural networks. In order to overcome the strong nonlinearities of N-value distribution for mountainous area, two types of new ideas are proposed. The one is to use an additional new feature value produced by using the geology type. The other is to use only learning data within a specific range, which is determined by using the proposed algorithm. The numerical simulation demonstrates the effectiveness of the proposed methods.
We aim to develop a swimming motion coaching support system for beginner and/or intermediate swimmers by using single inertial sensor. In this paper, we develop a prototype to visualize the acquired sensor data and display the timing information of swimming motion such as respective stroke and turn. Our previous proposed estimation methods of the swimming motion are implemented in the prototype. We carry out a questionnaire survey to evaluate the visibility, effectiveness and reliability of the prototype. As the result, although there are various opinions on the estimation error of respective stroke and turn, it is suggested that the prototype system is required at the actual training scene.
This study proposes a model in which the communication robot forms its personality based on its temperament and the other’s character. The proposed model learns how to behave in a group with Q-learning by considering its temperament. In our simulation, a robot using the proposed model and a pseudo-user interact with each other. Furthermore, the proposed model forms a different personality based on its temperament and the other’s character.
With the spread of CGM (Consumer Generated Media), a huge amount of digital data has been accumulated on the Internet. These data are utilized for improving social sensing technologies to measure not only social and economic trends but also various kinds of phenomena such as large-scale disasters. Using habitual behavior of users, authors proposed a new social sensing method for extracting phenomena in the actual world from the difference in the habitual behavior, and proved its usefulness. To practically evaluate the versatility of the application examples of the technology, it is necessary to clarify whether the data with different user attributes are also applicable. In this study, the habitual behavior of the users is analyzed attribute by attribute, and abnormal behavior is extracted based on different behavior from the normal time for each user attribute. Demonstration experiments were conducted to verify whether it is possible to find out social trends in the actual world or social phenomena for each user attribute on a detailed granularity.
With the development of communication devices including smartphones, utilization of SNS (Social Networking Services) such as microblogs is becoming more active. Microblogs contain posts about actual events such as news, and many methods for analyzing topics from these posts have been proposed. However, for approaches with a focus on posted content, when targeting microblogs in which new posts are added over time, the difficulty of structuring models that comprise topics is an issue. Therefore, this study develops a method of obtaining topics from the burst levels of multiple keywords with a focus on changes in accordance with time series in burst levels of keywords appearing in posted contents. In demonstration experiments, we conducted comparative experiments to compare existing topic extraction methods with our proposed method, and we verified that it is possible for our proposed method to detect topics that existing methods are unable to obtain.
Japan Sports Agency aims at supporting distinguished performance of national members of Japan from a scientific aspect in the prioritized policy concerning improvement in international athletic ability. Focusing on the field sports, we have been developing a system for visualizing athletes’ plays using the GNSS sensor. In particular, we have been performing a research on matchup analysis of pass plays with a focus on American football. The aim of the research was to decide whether a pass is completed or incompleted by deep learning, using trajectory images of matchup of QB, WR, and DB. However, since only the track information of the finished plays was used, it failed to obtain information perceived during the play, for example, prediction of the completed pass probability during the play, or information that enables directing timely choice or modification of action from the side line during the game based on the prediction values. In this research, we attempt to predict completion probability during plays by using track information that takes into account the time term. In experiments, we made analysis of quick passes, short passes, and long passes to demonstrate its usefulness.
In our country, policies regarding sports are actively advanced towards the Tokyo 2020 Olympic Games. One of those policies, ”Sports x ICT” considers effective methods of utilizing ICT (Information and Communication Technology), such as development of measurement instruments, measurement and visualization of data, and proposals for new services in the field of sports. Against this backdrop, we have been developing the visualization system for American football games using terminal devices included GNSS and acceleration sensor. Using that system, we confirmed to grasp the effective information that are the individual condition and the motion analysis of American football players against not only players but also the college football leaders. But, in our existing research, we could not realize the strategy analysis of games that are to select the play calls depending on circumstances, to make a prediction of successful ratio, and so on. Then, in this research, we apply deep learning to the matchup analysis of pass play included offence and defense players, and then verify whether it is possible to infer the success or failure of a play from it.
In recent years, accompanying the development of laser-based measuring technologies, initiatives relating to the utilization and application of three-dimensional information, such as point cloud data, have been gaining momentum. With focus on the field of roads, point cloud data has been widely used for surveying road surface conditions and detecting deformations in cylindrical features. Although it is necessary to identify the particular road features under analysis in these operations, point cloud data combines various types of road features together. Therefore, technology capable of classifying particular road features out of point cloud data is necessary. Existing research has proposed methods for extracting road features with high precision by separating road surface features from other data points. However, the problem is that it is difficult to accurately extract point cloud data related to the road surface when there are grade separations or successive slope changes because low elevation points are interpreted as the standard for the road surface. Thus, this research proposes a method for extracting point cloud data related to road surface features with high precision by estimating the elevation of roadways using plan of completion drawing.
In Japan, a large number of bridges constructed during the period of high economic growth are aging. The Ministry of Land, Infrastructure, Transport and Tourism has formulated a specification for creating a three-dimensional model and a definition of notation for the purpose of improving efficiency and upgrading of maintenance management of these bridges. In this background, a method was proposed for generating a three-dimensional model of a bridge using the point cloud data obtained from the laser scanner on the ground or UAV. However, in existing research, it is impossible to generate models for each individual part specified by the current specification. In this study, we propose a method for automatically recognizing individual parts of the main girder, the road surface, the handrail, and the abutment from the point cloud data of a bridge using deep learning.
Currently, the advent of the low birthrate and aging society in Japan has become a problem so it will be increased that the opportunity to resolve the problem by using robots at home. Therefore, it is convinced that robots will have more opportunities for being active at home. These days, we can get robots more inexpensively and it makes the communication between human and robot will be significant progress. When the robot is at home, the condition in which the robot can live together with residents is that the robot mimics residents’ experiences telling by words and gestures so that learns how or what to do at home.
The objective of this study is to make a robot enable to properly behave based on instructions given by people. We therefore consider a way of associating words with robot’s actions so that a robot can behave by understanding the meaning of words.
As a concrete example, we focus on various types of cooking actions represented by words with adverbial expressions and use multilayer perceptron to learn relation between adverbial expressions and robot’s actions. The meaning of elementary cooking instructions is represented with distributed semantics by means of word2vec.
To represent the actions of a robot, we have expanded the framework of Activity-Attribute Matrix (AAM) so as it can deal with the motion of actions.
We have employed multilayer perceptron to learn the correspondence between those actions and the meaning of the instructions, and confirmed how much the actions that a robot has never done can be precisely estimated with the meaning of the given unknown instructions with the learned model.
Robots are being expected to play the role of guessing the oblivion status in humans and intervening appropriately. This paper proposes an oblivion model concerning episode memory for communication agents. First, the subjects were instructed to listen to a certain sentence, immediately after which they were asked to freely describe the content. Moreover, one month after this event, the subjects were once again asked to describe the content, and their status of oblivion (quantity, quality) was investigated using the number of characters and idea units. The oblivion model for the communication agent was generated using the results of this investigation. Based on the results of the experiment it was presumed that, immediately after hearing the sentence it was appropriate to extract the content uniformly from the whole text, whereas after one month it was appropriate to extract the content centered around the main theme of the text. In this paper, the oblivion model corresponding to immediately after the reading (listening) was set as a text that had high score for surprise emotion of the communication agent. Moreover, the oblivion model after one month was set as a text that had the maximum score for the total of all the six emotions of the communication agent taking into account the peak-end rule. Additionally, the efficacy and issues of this model were verified through simulation applying this oblivion model on different texts, and comparison of the results with the oblivion status of the subjects in this study, and automatic summaries.
The ultimate purpose of this study is to explore the framework by which small and medium-sized companies are engaged in persona marketing. In particular, this paper shows the system based on an interactive genetic algorithm, with focus on pursuing the possibility that customers recognize the reality of Persona Design. The user only evaluates some scenarios generated by the system. Thus, it should be noted that a beginner without a specific knowledge can easily design the story of persona. We examine how 100 customers evaluate the reality of persona, which reveals the fact that our system is a valid tool to assume the real person.