The service industry is an important industry that accounts for about 70% of Japan's GDP. However, since the labor productivity of the service industry is lower than that of the manufacturing industry, productivity improvement in the service industry is the country's most important policy issue. In order to solve such problems, we research support method for sophisticated store management based on highly accurate future prediction for face-to-face service industry. As part of it, we research prediction methods using external data existing in the ubiquitous environment such as weather, events and internal data such as POS data etc. In this paper, we describe comparison of forecasting methods and material ordering for dishes based on machine learning.

The purpose of this research is to present a recipe suitable for the user’s cooking environment from various recipe candidates.Since recipes on user-contributed recipe sites are increasing yearly, it is not easy to find recipes suitable for his/her cooking environment such as cooking utensils and condiments that the user has. In particular, cooking utensils do not always described in the recipe texts. To solve the problem, a search system is desired that can retrieve cooking recipes by estimating omitted cooking utensils and condiments into account. As the first step of the system, this paper proposes a method to estimate omitted cooking utensils. This method uses foodstuffs and cooking behaviors appeared in the recipe texts as clues.As the result of applying this method, we confirmed that it is possible to infer the candidates for cooking utensil omitted in the recipe. This result shows the effectiveness of the proposed method.

The research related to gastronomy has been treated as a non-scientific one. The situation may change, through the development of artificial intelligence (AI). Dietary assessment is formed by combinations and interactions of multiple factors. This is not familiar with the so-called modern scientific method, trying to extract the effect of a specific factor. In order to draw out the potential possibilities of such AI, it is essential to develop a database on food culture.

In annual securities report, various information such as corporate policy, risk management, R&D, and so on, is included other than business performance. Previous researches proposed the extraction methods of important sentences containing causal information from financial articles and texts but not annual financial reports. In this paper, we applied these extracting methods based on SVM discriminant model to annual securities reports in our original way. Our method indicated high performance and all evaluations, that were precision, recall and F-score, showed more than 0.8. By using our model, useful information from annual securities reports would be collected effectively, which allow us to make unique investment decisions.

The literature review matrix is a method commonly used in writing a summary of literature surveys in the nursing field. In the literature review matrix, the literatures and literary groups obtained by the search are arranged on the vertical axis and the analysis aspects such as purpose, data, method, result and discussion are arranged on the horizontal axis, and features of each object are described in the cell. We applied machine learning to grasp the point of view being performed manually and to extract features of each view. We constructed a system that dynamically generates a review matrix according to the search. The present paper explains the idea, the method and case studies on 215 literatures on “text mining in nursing”.

We aim to develop a cooperative relationship between people and agents through interaction. To develop this, we have to make people recognize agents as intentional being, and make people interact with agents voluntarily. So we designed two agent models. The feature is to let people observe the situation, where two agents, the superior role and the subordinate role, interact with each other in the form of questions and answers. A psychological hurdle for interaction falls down and interacts voluntarily, by a man observes the state of making question and answering. I prepared experiment tasks to evaluate the performance of proposed agents. By the experiment, I proved that these proposed methods causes psychological attitude to feel the intention of the agent and voluntary interaction, and they are useful to develop cooperative relationship with agent.

The recent attacks in towns have increased demands for technology to find suspects in a crowd. A typical way to find suspects is to use facial recognition systems, however, those require a huge amount of personal data in advance. In this paper we propose a method to detect suspects in a crowd by observing pedestrians' behavior without using any specific personal data. The fundamental idea is that if a pedestrian finds a suspect, they might stop walking, or change the direction. We have devised a method to find such changes of behavior of pedestrians based on a Kalman filter and a hidden Markov model. The filter is used to detect a change, and the HMM is for assuming the intention of each observed pedestrian. Agent simulation results shows that the method works sufficiently well, especially where people are walking not in a single direction but in various different directions.

GPGPU is a parallel computation technology using GPU that has huge number of processor cores for parallelly calculating colors of pixels on a monitor. In a previous work, we used GPGPU to parallelize many runs of reinforcement learning agents for calculating their tness in a simulation of evolution. It speeded up the simulation surprisingly. However, the evolution part was sequentially run in CPU and the communication between CPU and GPU happened in every generation. Hence, this work uses GPGPU to parallelize the evolution part in addition to the tness calculation. It makes the simulation even faster due to parallelism and the reduction of latency between CPU and GPU.

We study hedonic games when each agent considers other agents as friends, enemies, or neutrals. In hedonic games, each agent has her preference over all coalitions that she can join. In this paper, we propose hedonic games under friends appreciation by allowing an agent to take into account the number of neutrals in her coalition. We show that the existence of neutrals does affect the stability of coalition structures, even though the impact of neutrals on preference is so small as to be meaningless compared to the impact of friends and enemies. More specifically, when each agent prefers coalitions with more neutrals, we show that neither a core stable coalition structure nor an individual table coalition structure can exist.

In this paper, we will propose method of estimating cyber threat intelligence with intelligent agents. Population of security operators are insufficient though cyber threat intelligence is increasing. Therefore, there is a possibility that the cyber threat intelligence can not be estimated properly. Our proposed intelligent agent is composed of three modules, Estimate module, Split module and Database. We will dene a problem of Estimate module and explain how to apply methods of machine-learning. And we will experiment with real data to verify feasibility of our proposed intelligent agent.

When a company goes bankruptcy, the effect can make another company go bankruptcy. This is called chain bankruptcy. In this research, we analyzed effect of asset price and inter-bank lending and borrowing network on chain bankruptcy using agent simulation. We have two results: (1) as the rate of change in asset price is higher, the final number of bankruptcy is higher; (2) as the density of links is higher, the final number of bankruptcy is lower. These results suggest that factor causing bankruptcy is both asset price fluctuation and inter-bank lending and borrowing network.

In real-time strategy games, it is difficult for a computer to defeat a human player under the condition that the computer has same recognition ability to the player. It is because many factors in the environment are changing in a very short time during which the computer has to choose an action. Although the Monte Carlo Tree Search (MCTS) algorithm obtains decent results in such games by searching a good action from simulation, it will be better if it has more time. On the other hand, it may be worse due to delay of response to the environment. In this work, we investigated the trade-off property in a ghting video game.

This paper presents a technique for exploring map or surface using extended L-Systems based on L-Systems’ grammar, other than L-Systems rules, we need this derivative rule for exploring a walkable area. The rule that used is flipping approach, which is a technique for changing the angle of generating grammar, then will impact the direction. The results are validated and done in simulation systems.

Basel regulatory framework, one of CAR (capital adequacy ratio) regulations, is said to make markets destabilized in a previous study. But the previous study included some inappropriate assumptions. So, this study assessed this destabilizing effects with a new model. In my model, FCN agents and 2 kinds of portfolio agents, CAR regulated ones and not regulated ones, were included. Using this model, some simulations were run. As results, the simulations revealed some facts: 1. Asset management using portfolio stabilizes markets and the stabilizing effect are significant if there are a lot of markets included in the portfolio; 2. CAR regulation destabilizes markets and vanish the stabilizing effects of portfolio. In addition, the results of my simulations suggest that CAR regulation does not only raise the chance of price crashes but also depress whole price.

The needs to employ machine learning is increasing for accurate estimation and noise reduction in recent advanced measurement where its output data is enormous, complex and noisy. Particularly, the recently emerging Positive and Unlabeled Classification (PUC) can be used to classify target objects and contaminants in the measurement. However, the existing standard machine learning is based on Bayesian estimation which assumes invariance of the target population distributions, whereas they are very different depending on the objects in the measurement. In this study, we investigated the PUC to overcome this issue. We applied the method to an actual measurement problem and confirmed its significant noise reduction.

Multi-label classification is a supervised learning problem where multiple labels may be assigned to each instance. The main baseline for multi-label classification is binary relevance method, which is estimate the binary classification model for each label. In binary classification, there are cases where poor results are data when the class is imbalance. In this paper, we propose a multi-label classification model used relative density ratio. In this model, we used relative F-measure by relative density ratio for weight of error function to solve the class imbalance problem.

In life-sciences domain, Linked Open Data (LOD) is being increasingly used when publishing databases. For flexible use of such databases, we employ a method that uses federated query search along a path of class-class relationships. To demonstrate our strategy, we implemented a prototype system accessible via a web API as a preliminary trial. We have been collecting SPARQL Builder Metadata (SBM) that describe a data schema for each SPARQL endpoint. Using the SBM, we constructed a graph called a merged class graph based on the class-class relationships in LOD. Using this graph, our system can provide the information required to construct a federated search query, such as the SPARQL endpoints that include a class–class relationship, and paths of the class-class relationships between two classes.

We studied personal attributes represented in tweets, such as gender, occupation, and age groups. First, we examined how much these basic attributes can be predicted from the texts of tweets, each of which was vectorized by a word2vec-based method for machine learning. The results showed that machine learning algorithms can predict all three attributes with 60-70% accuracy. We also confirmed that differences in word usage between males and females (related to semantic differences) affect the predictive accuracy of gender. Furthermore, we quantified other personal attributes, such as Big 5 and values, using IBM Personality Insights.

In recent years, a lot of non-verbal expressions have been used on social media. Ascii art (AA) is an expression by visual technique using characters. In this paper, we set up an experiment to classify AA pictures by using character features and image features. We try to clear which feature is more effective for the method to classify AA pictures. We proposed four methods; 1) character frequency based method, 2) character importance value based method and 3) image feature based method, 4) character's image feature based method. We trained the neural networks by using these four features. As the experimental result, the best classification accuracy was obtained with the feed forward neural networks using character's image feature.

Recent advances in artificial intelligence, especially deep learning, have enabled us to handle wider range of problems with computers. As for the real-world problem settings, however, there remain some difficulties, for example, inputs for embodied agents are partially observed representation of their states, and building models of their environments is needed for more sample efficient systems. One possible solution for coping with these difficulties is to use attention mechanism, which models the visual system of human and regards its inputs as a sequences, learning to where to attend. In this paper, we propose a method to train attention mechanism of neural network without external rewards. The proposed method consists of two ideas, one is to use intrinsic reward for attention mechanism and the other is to adopt adversarial learning in the model.

A new framework of a methodology of deterioration prediction using deep learning scheme for coastal protection facilities was proposed. As a basic examination, a deterioration prediction model was developed based on an inspection report of damaged facilities and observed data of wind-direction from 1966 to 1985. According to the simulation results, an average value of accuracy rate of the model was 0.44. Although sufficient accuracy could not be obtained in this trial, points to be considered for further investigations were discussed.

Super resolution, especially SRGAN, can generate photo-realistic images from downsampled images. However, it is difficult to super-resolve originally low resolution images taken many years ago. In this paper we focus on food domains because it’s useful for our service if we can create better looking super-resolved images without losing content information. Based on the observation that SRGAN learns how to restore realistic high-resolution images from downsampled ones, we propose two approaches. The first one is downsampling methods using noise injections in order to create desirable low-resolution images from high-resolution ones for model training. The second one is training models for each target domain: we use {beef, bread, chicken, poundcake} domains in our experiments. Comparing to existing methods, we find the proposed methods can generate more realistic super-resolved images through qualitative and quantitative experiments.

Deep neural networks (DNN) continuously demonstrates excellent performance in various application domains. However, how to control the representations is critical issues to use DNN in real-world scenarios. Notably, control the invariance of the representations is essential to incorporate social constraints, such as privacy-protection and fairness. This paper proposes a novel way to control the representations learned by DNN, called similarity confusion training. Empirical validations on a task of learning anonymous representations from the data of wearables show that the proposed method successfully remove unwanted information with less performance degradation compared to the existing methods.

The long-term care records keep the detailed information of daily life and behavior for the members in need of nursing care. In this paper, we conducted an analysis focused on the relationship between mimetic words in free description text of the records and nursing care level to demonstrate the real state and behavior of the people in each care level. As the results, text of milder level of care needed contains more active and positive type of mimetic words like don-don, wai-wai, whereas that of severer level of care needed contains dull and uncomfortable type of mimetic words like bon-yari. sowa-sowa. These results reveal that the mimetic words reflect the differences of people’s behavior in each care level. Further analysis focused on the mimetic words toward the massive text data should be effective to excavate the collective wisdom from long-term records.

The authors have been studying haiku generation by computer. A haiku composed of fragmentary words with, basically, only 17 syllables and haiku generation by computer is an interesting theme. Although there have been various approaches to haiku generation, in this paper, based on our previous haiku generation studies including the following two approaches or methods, the authors present a consideration on the possibility of a unified approach of the top-down generation methods using mainly symbol processing techniques and the bottom-up generation methods using chiefly neural processing such as deep learning techniques.

Recently we have been discussing how to express the taste of Japanese sake. During describing the taste of Japanese sake, we found a question whether the expression of the taste is abstract or representational. Previously, I discussed the effect of drawing the taste. The drawings can be regarded both abstract and representational. Because they were drawn because it was quite difficult to express the taste by words. Accordingly they may be abstract. However, if we chenge the viewpoint, it can also be regarded representational. Because they described the taete which we could not see.

Taste of Japanese sake is frequently expressed with the expression ``Karakuchi (dry taste).'' However, there are many kinds of taste and flavor in Japanese sake. We have analyzed monthly sake reports delivered from Nishimura sake shop at Kyoto, which include various kinds of expressions of sake taste such as flavor, sweet taste, bitter taste, sour taste and so on. We tried to analyze similarity between sakes in terms of several points of taste expression using syntax analyzer CaboCha. As the results, some similar taste Sakes were extracted based on this method, but it will be necessary to use more information such as word similarity of Sake taste expressions.

In this study, we implemented a gaze model, based on human gaze behavior analysis, on an eyeball manipulable robot and CG agent. Then we analyzed the effects of agent embodiment and change in gaze amount on agents' personality perception. The results suggested that it is possible to express extrovertedness and confidence of the agent by changing the gaze amount from the agent regardless of agent's embodiment. In addition, the levels of perceived extrovertedness as increase in gaze amount differ depending on the embodiment of the agents. The CG agent's perceived extrovertedness increased in proportion to increase in its gaze amount, while the robot's extrovertedness increased logarithmically to increase in its gaze amount.

Active listening is one of the indispensable axes in evaluating the dialogue of psychotherapy. Although there have been discussions about it in the area of clinical psychology, the method for evaluating active listening has been missing. It is thus necessary to establish it to improve the quality of listening in the interview. The authors have proposed a measurement method of the degree of active listening with a device we originally developed to evaluate emotion (EMO system). This paper reports on the experiment conducted to compare the evaluations of a psychotherapy by expert clinical psychologists with those by undergraduate students as one of the coursework tasks for a clinical psychology course. A new experimental setup was proposed including a multiresolutional analysis to detect the change of active listening evaluation.

Automatic meeting summarization would reduce the cost of producing minutes during or after a meeting. With the goal of establishing a method for extractive meeting summarization, we propose a multimodal fusion model that identifies the important utterances that should be included in meeting extracts of group discussions. The proposed multimodal model fuses audio, visual, motion, and linguistic unimodal models that are trained by employing a convolutional neural network approach. The performance of the verbal and nonverbal fusion model presented an F-measure of 0.827. We also discuss the characteristics of verbal and nonverbal models and demonstrate that they complement each other.