Vol.2 No.1 2009

8/88

Research paper : Predictive modeling of everyday behavior from large-scale data (Y. Motomura)−5−Synthesiology - English edition Vol.2 No.1 (2009) Further, by using a probabilistic model, it is also possible to model abstract and diverse elements, including, for example, the idiosyncrasies of individuals, making it possible for conventional psychology to attempt to deal with a universal human model. This is an important viewpoint for models that implement individual adaptation or personalization crucial in information processing associated with recent human centered design and usability aspects. There have been various implementations of user modeling which use a Bayesian network as the probabilistic model [6]. In order to model, in particular, constructions of human cognition and evaluation as Bayesian networks, interview methods used in clinical psychology or marketing are applied [7]. In this way it is possible to infer acceptability or intention by modeling and implementing probabilistic inferences of the system and services of the user.5 Modeling of Everyday ActivitiesAs an actual example of computing uncertainty, the standpoint of user modeling has been discussed above; however, when considering everyday life assistance as various actual services[8], modeling of the living person, each day, as a user, is vital. Until now, sensors have been installed in a home, or a “sensor house” has been proposed for research and development in order to analyze everyday activities [9]. Until now, several applications have been proposed wherein abnormalities are judged by detecting outlying values while modeling patterns in measured data as stationary distributions; however, for broader applications, modeling with only stationary distributions is inadequate. Rather, it is necessary to consider optimizing the utility, or value, depending upon user intentions. In other words, in order to predict the intentions, impressions, and assessments of the user, which cannot be observed directly, from observable actions, higher-order inference is necessary. To accomplish this it is necessary to model dependencies and causal relationships, such as how results turn out in response to certain circumstances and activities; to accomplish this, it is necessary to record, not only action data, but comprehensive data involving causal variables, and to search for causal structure from relationships among a large number of variables.This can be considered as a new kind of analysis of behavior, opened up by sensor technology and modeling technology. Behavior analysis was established by Skinner in the mid 1900s as a field of research making use of the behavioral science approach within psychology [10]. In this context, human activities are referred to as antecedent and behavior contingencies, and expected changes in environment, as the result of actions, are thought to be determined from relationships among three items. Further, the causal relationships between antecedents and behavior contingencies are clarified when focus is on a certain activity. While modeling this explicitly and by causing changes in those behavior contingencies and antecedents, control of behavior is implemented.It is necessary to interpret video images of observed activities and to perform a labeling procedure in order to discover the cause and effect of activities. Performing this manually, however, requires an enormous amount of time and effort; therefore, it is difficult to efficiently analyze natural activities in an everyday life environment. In addition, performing this interpretation manually allows only a small number of objects to be analyzed as control variables of the activities. Technology capable of automatically handling large amounts of observational data is necessary in order to analyze everyday activities.In such cases, observing actions automatically by a sensor network embedded in the environment, and utilizing statistical study techniques comes to mind. By constructing a Bayesian network model through statistical learning from the large amount of sensor data gathered, it is possible to connect the inevitability amid the reasons and purposes for actions, which become candidate behavior contingencies, and the environment and situations, which become antecedents. In this way, it is expected that behavior analysis will be developed largely by contributions from model construction technology that extracts variables with strong causal relationships from sensor technology that can comprehensively observe daily activities, and from the data observed thereby. Through modeling of actions based on Bayesian networks and ultrasonic sensor networks, research has, until now, been aimed at analysis of everyday life behavior [11] and at applications such as injury prevention for children [12][13]. An example of inferring the behavior of children [14] is introduced below.Attach an ultrasonic transmitter to a person or object in the room. Then, at regular intervals, position information (x, y, z coordinate data) of the person or the objects can be captured by ultrasonic receivers embedded in the sensor room. A fisheye camera (camera with a wide-angle lens) installed on the ceiling of the room simultaneously photographs the circumstances of the person’s activity in the room as a video. For the activities of the person in this photographed room, the video images are labeled manually at one-second intervals. For example, a detailed database is collected giving action labels such as “the person is walking,” “sitting,” “standing.” Modeling of everyday activities is performed using this data, and experiments of behavior inference are performed based on the video.Considering the problem as the system’s observation of the behavior by means of sensors and images, it can be formulated as a type of pattern recognition problem. Since data arising in the everyday real world considers human life activities and the living environment as a background, the nature of the state

※このページを正しく表示するにはFlashPlayer9以上が必要です