人工知能学会論文誌 Vol. 32(2017) No. 5

The growth of robot technology has prompted growing interest in educational-support robots that assist in learning. Most of these studies report on collaborative learning between educational-support robots and healthy children. Meanwhile, the number of children in primary schools with diagnosed developmental disabilities (gray zone children) has increased in Japan. Gray zone children may have difficulty learning over long time periods. Moreover, gray zone children tend to receive peer teaching from healthy children in the school environment. Other symptoms of autism in children are low self-esteem and possibly depression. We expect that gray zone children will learn best by teaching another learner. Learning-by-teaching promotes self-esteem and improves the learning time. In a previous study, a robot that answered a question incorrectly and uttered “Please teach me” or similar statements provided a collaborative learning environment for the learning-by-teaching method. However, whether collaborative learning with this robot increases the learning time of gray zone children was not investigated. Therefore, the present study investigates whether gray zone children can improve their learning time in collaborative learning with a robot that prompts learning-by-teaching. The robot is designed to answer questions incorrectly and utter statements such as “Please teach me.” The robot is also designed to have learning capability. For example, the robot learns the methods of problem-solving from its human partner. Thus, when presented with a question that can be solved by a previously learned method, the robot can answer the question correctly. The experimental results suggested that the learning enhancement was driven by the robot’s initial incapacity to answer a question, and its requests for assistance by the gray zone child. Gray zone children engaged in collaborative learning with our robot spent more time learning than those working alone. Moreover, the gray zone children enjoyed the collaborative learning with our robot than the robot which always solves questions correctly and never solves questions correctly.

In order to avoid the risk of bankruptcies’ succession in financial institutions, research on financial transaction networks on systemic risks has been active globally, mainly in Europe. In this research, we propose a new strategy to reduce the bankruptcies’ succession at minimum cost by constructing an inter-bank transactional network model composed of Erdos-Renyi network and Barabasi-Albert Model considering network property. As a result of the verification using agent-based modeling, it is verified that financial assistance implemented to stop the bankruptcies’succession will eventually increase the succession, andWe clarified the importance of selection of financial institution implementing financial injection. We also show that eXtreme Gradient Boosting is effective as a method of selecting financial institutions and propose new strategies to improve accuracy.This is a major economic effect in the sense that it proposes the risk indicator that aims to improve the certainty of the investment of public money by analyzing the properties of the transaction network analysis of financial institutions.

The growth of robot technology has prompted growing interest in educational-support robots that assist in learning. We focus on a robot that prompted learners to alternately solve questions in collaborative learning. The previous study reported that robots that alternate between the roles of the speaker and listener can prompt college students to learn by alternately solving questions with the robot, possibly achieving the same effect as collaborative learning between two college students. As the speaker, the robot explained a solution to its partner and solved a question. Moreover, the robot improved its accuracy rate as learning progressed. As the listener, the robot did not solve questions itself, but instead payed attention to its partner that solves the questions. However, the previous study did not investigate the method by which the accuracy rate of the speaker robot was improved. Therefore, this study investigates the manner in which change in the accuracy rate in speaker robots prompts effective collaborative learning. Herein, we define collaborative learning as learning in which students and robots alternatively solve questions. We compare four methods of changing the speaker robot ’s accuracy rate in the experiment. In the first method, the accuracy rate changes in a manner similar to the learner ’s accuracy rate. In the second method, the accuracy rate is initially set to 0% but is gradually increased (in 10% increments) to 100% as learning progresses. In the third method, the accuracy rate is set to 100%; therefore, the robot always solves questions correctly. In the fourth method, the accuracy rate is set to 0%; therefore, the robot never solves questions correctly. The results of this study suggest that there was no difference in the learning effect of each group. However, we found that the robot that improved its accuracy rate as learning progressed could prompt learners to feel greater friendship for it than the robot which always solves questions correctly. Moreover, the learners could alternately solve questions with the robot that improved its accuracy rate as learning progressed more effectively than the robots of other groups. Therefore, we believe that a robot alternately solving questions with a human while constantly improving its accuracy rate may be the best for collaborative learning with a human.

This paper presents a novel metric for evaluating stability of machine translation system. A stable system indicates that it keeps almost the same outputs given the inputs with slight changes. In this paper, we propose a stability metric by exploiting TER metric for evaluating the differences between the two texts. We have built an evaluation data set, and demonstrate that a neural-based method is unstable rather than a statistical-based method, while the former outperforms the latter.

Keeping a living and working spaces tidy is very important for healthy daily life. The efficiency in working rises when the office is tidy. We developed a system that tidies up through the cooperation between a robot and a human. In this study, we investigated the influence of a robot’s behavior on the motivation of tidying up. For completing this system, it is necessary to investigate the effective behaviors that encourages a human. To validate what behavior effectively encourage human to tidy up, we conducted a preliminary experiment with 13 male and 5 female participants, aged 20-23. We found the statistically significant difference between the cases where the robot took actions or not.

Facial expressions play an important role in communication as much as words. In facial expression recognition by human, it is difficult to uniquely judge, because facial expression has the sway of recognition by individual difference and subjective recognition. Therefore, it is difficult to evaluate the reliability of the result from recognition accuracy alone, and the analysis for explaining the result and feature learned by Convolutional Neural Networks (CNN) will be considered important. In this study, we carried out the facial expression recognition from facial expression images using CNN. In addition, we analysed CNN for understanding learned features and prediction results. Emotions we focused on are “happiness”, “sadness”, “surprise”, “anger”, “disgust”, “fear” and “neutral”. As a result, using 32286 facial expression images, have obtained an emotion recognition score of about 57%; for two emotions (Happiness, Surprise) the recognition score exceeded 70%, but Anger and Fear was less than 50%. In the analysis of CNN, we focused on the learning process, input and intermediate layer. Analysis of the learning progress confirmed that increased data can be recognised in the following order “happiness”, “surprise”, “neutral”, “anger”, “disgust”, “sadness” and “fear”. From the analysis result of the input and intermediate layer, we confirmed that the feature of the eyes and mouth strongly influence the facial expression recognition, and intermediate layer neurons had active patterns corresponding to facial expressions, and also these activate patterns do not respond to partial features of facial expressions. From these results, we concluded that CNN has learned the partial features of eyes and mouth from input, and recognise the facial expression using hidden layer units having the area corresponding to each facial expression.

Cities are regarded to complex systems that are explained by various components and their interactions. Today, the more complex the components and interactions become, the harder to solve urban problems become. Due to their complexity, it is very hard to forecast the effect of urban policies and make effective ones as well. Sometimes policies are less effective than expected and do not lead to solving urban problems. If the natural and social dynamics by simulation models can be described well, we can assess and verify urban policies to solve various urban problems. Although there have already been such researches using macroscopic or microscopic models, in the former models the interactions between city components are not considered enough and in the latter models there are problems in showing their validity and data availability. Therefore, some researchers aim to integrate two or more models, each of which describe just one urban phenomena. Given that integrated models are consist of various simple component models, they are useful and suitable to evaluate urban policies. Now we think that households are quite important among such city components as decision makers. Our goal is to develop an integrated urban model including a land-use model, a finance model, a household model and so on. In this research, we propose a multi-agent-based household transition model as a basis of such integrated model. Our approach is to use a mesoscopic model that makes us to consider individuals and households. We validate our model through two simulations about Kanazawa and Yokohama, using their real statistic data.

Government 2.0 activities have become attractive and popular these days. Using tools of their activities, anyone can report issues or complaints in a city on the Web with their photographs and geographical information, and share their information with other people. On the other hand, unlike telephone calls, the concreteness of a report depends on its reporter. Thus, the actual status and demand to the status may not be described clearly or either one may be miss-described in the report. It may accordingly happen that officials in the city management section can not grasp the actual status or demand to the status of the report. To solve the problems, automatic finding incomplete reports and completing missing information are indispensable. In this paper, we propose methods to detect parts related to an actual status or demand to the status in a report using empirical patterns, dependency relations, and several machine learning techniques. Experimental results show that an average F-score and an average accuracy score our methods achieved were 0.798 and 0.893, respectively. In addition, in our methods, RF achieved better results than SVM for both F-score and accuracy scores.

In recent financial market, high frequency traders (HFTs) and dark pools have been increasing their share. Financial analysts have speculated that they might decrease market transparency and malfunction price discovery, and their interaction would make the situation worse.To validate speculations, artificial market simulation is a tool of study by constructing virtual markets on computers. In this research, by constructing an artificial market simulation, we analyzed how the interaction between HFTs and a dark pool impacts on the market efficiency (in the sense of price discovery) of a (lit) stock market. In simulations, two types of trader agents enter the market. A market maker agent, a representative strategy of HFTs, submit orders to the lit market. We analyzed the market maker's interest rate spread, or simply the spread, as a key parameter for their strategy. Stylized trader agents submit orders to either the lit market or the dark pool with some probability given as a parameter.The simulation results suggest that on the condition that market makers have little impact to market pricing (having a large spread), moderate use of dark pools can promote market pricing. On the other hand, on the condition that market makers have big impact to market pricing, excessive use of dark pools can inhibit market pricing, while using dark pools do not have bad influence when the rate of use is not high. On the influence of market makers, our results suggest that the bigger the impact to market pricing (a small spread), the more it can promote market pricing.

Crowd simulation is one of the most widely used technique for the design and evaluation of the human-in-theloop situation such as evacuation plans, the building designs and so on, in a virtual environment. In order to have a valid evaluation, it is necessary to have a correct model of an individual agent’s decision process which causes a behavior of human’s in a crowd. However, in general, designing a decision process of agent’s largely depends on a trial-and-error manner. To avoid the trial-and-error by human designers, we focus on the automated method to derive agent’s decision strategy from the real data of human’s. In this paper, we consider a decision process consists of two stages. One is the strategy phase to select a goal state, and the other is the policy phase to output the primitive action of agent’s. We focus on the strategy phase. Though it should be more natural to assume that the strategy of each agent is not all the same, the existing method assumes that all agents have a common and homogeneous strategy. The proposed method makes it possible to extract the individual and different strategies of agent’s by evolutionary computation. The results of the experiments show the validity of our method. In addition, it is shown that there exist the cases where multiple strategies will be extracted for a single trajectory.

The cake cutting problem is concerned with the fair allocation of a divisible good among agents whose preferences vary over it. Recently, designing strategy-proof (SP) cake cutting mechanisms has caught considerable attention from AI and MAS researchers. Previous works assumed that an agent’s utility function is additive so that theoretical analysis becomes tractable. However, in practice, agents have non-additive utility over a resource. In this paper, we consider the all-or-nothing utility function as a representative example of non-additive utility because it can widely cover agents’ preferences for such real-world resources as the usage of meeting rooms, time slots for computational resources, bandwidth usage, and so on. We first show the incompatibility between envy-freeness (EF) and Pareto efficiency (PE) when each agent has all-or-nothing utility. We next propose a SP mechanism that satisfy PE, which is based on the serial dictatorship mechanism, at the sacrifice of EF. To address computational feasibility, we propose a heuristic-based allocation algorithm to find a near-optimal allocation in time polynomial in the number of agents, since the problem of finding a PE allocation is NP-hard. As another approach that abandons PE, we develop an EF and SP mechanism. Furthermore, we argue about false-name-proofness (FNP), which is the expansion of SP, and propose FNP and EF cake cutting mechanism. Finally, we evaluate our proposed mechanisms by computational experiments.

In this research, we show a paradox of the “theme park” problem. In the crowded amusement park, it is generally believed that the equalization of queue lines of people can decrease the waiting time for riding on attraction. However, the equalization of queue lines occasionally increases the waiting time in the case where congestion degree is over the limit of capacity. This paradox makes it difficult to reduce congestion. In this paper, we propose a method to reduce the waiting time even in the “theme park paradox” situation, and evaluate effectiveness of our method by multiagent simulation.

Evacuation planning is important to mitigate the ill effects of a disaster, such as a fire in earthquakes. For the evacuation of pedestrians, the route choice should maximize the completion rate of the evacuation. Some models of route choice have assumed that pedestrians will recognize the road conditions and the shortest route to the refuge perfectly. However, the validity of the assumption is controversial. In this paper, we propose a new model of route choice, which considers the differences in map recognition between individual pedestrians: the position of the refuge, the cognition of the road and other factors. Then, we discuss an evacuation of pedestrians from a fire, based on the model, including changing the pedestrian’s recognition of the factors. We also utilize a microscopic pedestrian model for simulating the behavior of the pedestrian in the continuum space, based on its visibility. For example, the recognition of the route speeds up the evacuation and raises the completion rate of the evacuation, however, the effect is slight. In contrast, the pedestrian’s recognition of the refuge position more significantly affects the completion rate of the evacuation. These results imply that even rough guidance, such as giving pedestrians the direction of the destination, could increase the completion rate of evacuations significantly.

In recent years, pedestrian flow modeling, simulation and optimization has received a surge in attention. There are two reasons for investigating pedestrian flow. One is to solve the day-to-day problems of crowded public spaces, such as stations, shopping malls, pavements and many other places. The other is to optimize evacuation procedures from densely occupied buildings and urban centers in case of an unpredicted event. Such evacuation planning can save many lives at sudden catastrophic events like tsunami. As walking is the most basic mode of transportation, many pedestrian models have been proposed. However, finding a realistic description of pedestrian behavior is extremely challenging and pedestrian modeling still remains a matter of debate. For this, the elaboration of a model is necessary. This research proposes a microscopic pedestrian simulation model based on concepts of cognitive science. In this model, the pedestrian utilizes visual cues as primary source of information to decide where to walk. The pedestrian movement is generated from bottom-up by describing behavioral heuristics and motion control process. It is clearly demonstrated by numerical analyses that this model reproduces human-like individual trajectories and facilitates smooth pedestrian traffic flows through the effective avoidance maneuvers. Our approach has potential for contribution toward investigating disaster risk reduction strategies in complex urban space.

In recent years, there has been a growing demand for low carbon and energy saving in the transportation sector. Meanwhile, some auto mobile companies started to introduce their own electric vehicles (EVs) better environmental performance than gasoline vehicles into the Japan market in 2009. However, their widespread adoption is hindered by a plurality of different factors in comparison with gasoline vehicles, i.e. short cursing distance, expensive purchase price, lack of charging stations and long charging time. Various researches and developments have been actively conducted in order to solve these problems. Among them, recommendation of energy efficient path via charging station (CS) is one of the promising solutions. In this paper, we propose a new path finding algorithm in consideration of EVs’ charging behavior on microscopic traffic simulation. The proposed method was validated through comparison experiments on grid network environment with ad hoc methods in conventional research. As a result, it showed that the proposed method yields an energy efficient route although the computation time increases. In addition, it shows good robustness under congestion and fluctuation of traffic volume, as a result of comparing the probability of out-of-charge and the amount of power consumption in the numerical experiments between proposed method and ad hoc method.