Vol.2 No.1 2009

4/88

Research paper−1−Synthesiology - English edition Vol.2 No.1 pp.1-12 (Jun. 2009) control the system based on decision theory, express useful knowledge, and perform complex processing, calculations with high-dimensional probability distributions involving multiple variables are necessary. As the number of variables becomes enormous, calculations involving high-dimensional probability distributions become complex; therefore, one has no recourse but to approximate locally using low-dimensional probability distributions. In order to facilitate this, a graph structure is introduced which stipulates the relationship between variables. As a multidimensional probability distribution model having this type of graph structure, we have the example of a Bayesian network [2]. Bayesian networks are general models that stipulate dependencies among many variables by a conditional probability and network structure. Bayesian networks can construct a model by statistical learning from large-scale data, which in turn becomes an important feature in handling uncertainty.In the current work, after discussing the non-deterministic approach and probabilistic modeling, together with Bayesian networks and techniques of constructing models that use them for predicting human activities in everyday life, actual cases in which they are applied are discussed. Finally, a hypothesis about “Research as a Service,” the construction of which has become inevitable in the process of implementation, is proposed and discussed.2 Selection by Non-deterministic ApproachIn real-world problems, we want to know the situation (value) or the possibility (probability) of objects that cannot be observed directly (latent variables). This type of uncertainty inevitably enters into computing when humans are considered as the object. When a system implements any task, the tasks are modeled within the system and considered to be the object 1 IntroductionThe range of applications of information processing technology is steadily increasing. At the same time, information services to aid in everyday life are increasingly in demand. Therefore, a model is necessary which describes the activities of daily life in a quantifiable way, in terms of what a person is trying to accomplish in various circumstances. Using such a quantitative theoretical model, we consider a system that predicts background requirements and expected results from the user’s activities, rapidly implements them, and makes possible the development of new services that aid activities of everyday life. Additionally, by continuously implementing such cooperative operations with people during everyday life, it becomes possible to acquire meaningful data in large quantities not previously obtainable in a laboratory environment. Using this large-scale data, it is possible to bring about a cycle in which services continue to be used while the model is constantly being updated. However, during these daily activities, information processing based on uncertain information (such as predictions that result in indeterminate or incomplete observational information) is of fundamental importance. What is needed is a paradigm shift from the deterministic approach, which has until now played a central role in system recording methods, to a non-deterministic approach. The non-deterministic approach is an approach to calculation in which ambiguous or uncertain information is processed as is, as far as possible. Calculations are made with the probability distribution as an object variable, along with the stochastic inference, which makes the prediction [1]. This stochastic inference has come to be used naturally as a naïve Bayes model or a Hidden Markov Model (HHM) using, for example, a pattern recognition device that maximizes the posterior probability. Further, in order to - Learning and inference from Bayesian networks based on actual services -Yoichi MotomuraCenter for Service Research/Digital Human Research Center, AIST Tsukuba Central 2, Umezono 1-1-1, Tsukuba 305-8568, Japan/Aomi 2-41-6, koto-ku 135-0064, Japan E-mail : y.motomura@aist.go.jpReceived original manuscript September 24, 2008, Revisions received January 13, 2009, Accepted January 13, 2009Daily life behavior modeling is discussed. This modeling framework consists of statistical learning, probabilistic reasoning, user modeling, and large-scale data collecting technologies. Bayesian networks can represent causality relationship as graphical structures. Such models should include situations and contexts of daily life behavior through real services. In order to collect large-scale data connected with them, we have to provide real services supported by many users. This concept is named “Research as a service” and discussed in this paper.Predictive modeling of everyday behavior from large-scale dataKeywords : Bayesian network, statistical learning, probabilistic reasoning, user model, behavior analysis, knowledge circulation[Translation from Synthesiology, Vol.2, No.1, p.1-11 (2009)]

※このページを正しく表示するにはFlashPlayer9以上が必要です