User type estimation is an integral problem in providing products and services. One of the most effective ways to estimate user types is to request users to complete a questionnaire. In most cases, a questionnaire has multiple questions and requires considerable user time to complete. To lower the cost of evaluating user type based on a questionnaire, we need to reduce the number of questions in the questionnaire. In our previous study, we proposed an adaptive questionnaire system based on a Bayesian network, which sequentially presents questions according to respondents' answers and then estimates the respondents' user types. To improve the accuracy of respondents' type estimation, we propose two modification methods for this system. One method is the reconstruction of the Bayesian network to reduce the number of questions displayed. The other method is the change of the number of respondents' type estimation. Experimental results show that the latter method is superior to both the former modification method and the original method.
User behavior modeling becomes important to create user-friendly products and services. To understand user behavior, we have to know the users' thinking, feelings, and "kansei." In our previous studies, we took a questionnaire to car purchasing users and made probabilistic user models about car selection. In these analyses, we showed that the selection of cars depends not only on the life-stages of customers but also on their kansei. In this study, to expand utilization of user modeling, we make an information recommendation system based on a kansei user model. This system recommends information about premium beer according to a respondent's philosophy concerning beer drinking and attitude to life, and then it measures the satisfaction level with regard to the recommendation. We tested this system at a museum having a history of beer exhibition and offered new beer drinking space. We showed that most users were satisfied with the recommendation results. We also showed that a number of museum users seemed to have little interest in beer and beer drinking.
In theWerewolf game, players's votes greatly affects winning and losing. We propose a method to analyze utterances of players by using Bayesian Networks and decide voting destination. There are numerous factors that determine the voting destination. Therefore we identify the factor that strongly influences the decision of the voting destination by use of decision tree analysis. In addition, a player's program is create based on the inference model constructed by the proposed method and its effectivenesss is verified.
As the number of reported child abuse cases is increasing, the workload of child welfare social workers is highly escalated. This study aims to find the characteristics of recurrent cases in order to support the social workers. We collected data around the child abuse and neglect from a prefecture database and analyzed it with Probabilistic Latent Semantic Analysis and Bayesian Network modeling. As the result, pLSA showed the four different clusters and Bayesian Network revealed a graphical model about the features of recurrence cases. The Interpretable modeling can be effectively deployed in those child welfare agencies to save children who are suffering from child abuse cases.
Due to the development of the sensor technology, the driver assistance system and autonomous driving system are intensively studied. Prediction model is then required based on the driving behavior. In this study, we construct the driver model to predict the behavior at a crossroad, using the brake timing sampled from the event data recorder/ blax box. The density estimation is the key technique since the moment of the braking is a type of data observed as an event not as a continuous one such as the velocity. Thanks to the smoothing of the occurrence points of the event, we can calculate the probability at any place along the route, which enables us to establish a method to predict the route from the braking behavior.
In this paper, we consider the validity of ranking of agents in RoboCupRescue Simulation. The simulation contains stochastic variable. Therefore, rank the agents by simple ranking methods such as the average score of multiple experiments. We de ne a proper ranking and discuss differences between it and simple rankings through experiments.