Vol.2 No.1 2009

5/88

Research paper : Predictive modeling of everyday behavior from large-scale data (Y. Motomura)−2−Synthesiology - English edition Vol.2 No.1 (2009) of calculation operations. In other words, the program can be understood to be the model and computing operations of the object tasks, coded by means of a programming language. Further, the same operation cannot be performed for all users; rather, it is necessary to invoke both a model of the task and a model of the user within the system. Ordinarily, models of the task (process) can often be given clear descriptions; however, for users, in order to treat uncertainties associated with humans, a non-deterministic analytical model is often necessary. In the current state of affairs, clearly modeling the latent variables of human relations, such as the intentions or requests of users, is difficult; and one is compelled, in the first place, to describe this type of situation using a non-deterministic framework. In addition, when a variety of users utilize the system in various circumstances, stipulating beforehand all of the most appropriate operations the system should take is a difficult problem. The system designer should design in advance the capabilities offered by the system; however, answers to questions such as what the system user requests, how the user will react to information or services offered, were the system operations correct, were the user expectations met, etc., will not be known until the system is executed or even after it is executed. In other words, it is difficult to decide, in advance, the most appropriate design of operations for users. Consequently, merely using a non-deterministic framework to operate the system as requested by, or as expected by, users is inadequate; and having predicted user reaction at execution time, frameworks allowing dynamic construction of user models become very important in optimizing these reactions and evaluations. This is the uncertainty associated with humans.Information to be calculated appears in large amounts, and large gaps emerge between computable amounts. However, handling uncertainty is necessary. For example, through the spread of the Internet, we find that limitations exist; and handling it directly necessitates facing an unwieldy amount of data. It is possible to calculate the frequency with which a given web page is read by all users; however, deterministic processing, such as counting this for all web pages, is not realistic. In this case, Google’s PageRank is calculated by modeling the transition probability among web pages non-deterministically as a stationary stochastic process [3]. In other words, we describe deterministically the construction of source pages or related links, and although this is a computer-based or deterministic method, it is not contained within a deterministic framework. Being a strategy that uses a non-deterministic model, it can respond to an explosion in description quality or the number of data points. Coping with uncertainty in a system involving this type of real world or large-scale data that includes humans is highly desirable in an artificial intelligence system, in order to tackle real societal problems; and here, describing problems using a non-deterministic model is one solution.Even if problems involving uncertainty are described by a non-deterministic computational model, the current deterministic computer processing brings to mind the computational theory of Marr [4] in which the following types of questions are considered independently: what is to be computed, how does the calculation method write the computational process (algorithm level), how is it to be implemented, etc. In other words, even if implemented by a program described by a deterministic computer language on a deterministic silicon chip computer, and, as in the previous web example, even if the original data or mechanism is deterministic, it can be profitable to think of the model as being calculated non-deterministically. As far as “toy problems” are concerned, it is sufficient to consider calculated quantities deterministically; however, one cannot avoid using a description with a non-deterministic framework when attempting to computationally model actual pressing problems in order to cope with the uncertainty existing within them.3 Bayesian Networks3.1 Probabilistic ModelingAs one non-deterministic approach, there are probabilistic methods. By using probability, it becomes possible to quantitatively model the non-determinism of phenomena and to treat it strictly by means of axiomatic probability theory. The probability values to be assigned to observable phenomena can be obtained from a large quantity of observational data; and for unobservable phenomena, estimates may be made by Bayesian probability theory (Bayesian hypothesis). This models the uncertainty of variables and the relationships between them through conditional probability and is easy to understand when considered as a way of determining the uncertainty of a particular variable, given information about other variables. Since this unknown probability distribution is treated as a subjective prior distribution in the conventional Bayesian hypothesis, it has been criticized by non-Bayesian statisticians. Recently, however, as large amounts of data have become more manageable, it has become possible to empirically construct this probability distribution from large statistical data sets and this approach is promising as a practical method in domains having a large amount of uncertainty.For example, consider a stochastic framework for treating totally unobservable phenomena. There is a large amount of uncertain information in the real world, such as future weather conditions, noise signals, or user intentions, for which it is difficult to determine a specified value. We introduce a stochastic framework in order to systematically cope with these. An object that includes indefiniteness, such as complex factors or the influence of noise, will be represented as a random variable denoted by X, and the concrete values this variable can take will be represented by x1, x2, …, xn.Next, consider dependencies between variables. For

※このページを正しく表示するにはFlashPlayer9以上が必要です