For time series data, we smoothly approximate its linear chart by a linear combination of Meyer wavelet functions. We define the error function as the H between the approximating function and the function which corresponds to the linear chart. An approximating function is called optimal if it minimizes the error function. To seek the optimal approximating function, we use simulated annealing. We call this method WISAM (Wavelet Interpolation with Simulated Annealing Method). In general, the numerical calculation of simulated annealing becomes a simulation of a certain timely inhomogeneous Markov chain, which is called an annealing process. In our case, the annealing process can be said to be a modified simple random walk. From optimal approximating functions obtained for sample data, it is suggested that the value of the error function for an optimal approximating function can be a measure of the magnitude, of the time variation of the data. As an application of this results, we propose a method of analizing the variation of the time series data and examine it to stock price data.

View full abstract

Construction of experimental design is one of the most important fields in statistics. Balanced incomplete block design(BIBD), one of the most useful designs, is available to save the number of experiments. However, in general, it is not easy to construct a BIBD since we need theory of finite field, finite projective geometry, Latin square and so on. Feed-Forward neural networks are applied in many statistical data analysis such as regression analysis, discriminant analysis and so on. However, Hopfield neural network which has feedback effects is not taken notice though it is shown to be useful in traveling salesman problem and other many optimal problems. In this paper, we consider a method to construct a BIBD. The method is based on the numerical computation of a dynamical system induced from a Hopfield neural network. It is shown that the method is successful by trial and error method with random initial points.

View full abstract

In this article, we discuss methods to construct confidence bands for regression function, especially to construct uniform width-confidence bands and trapezoidal confidence bands in a finite interval of interest. With case studies and numerical experiments, we examined the coverage probability and the efficiency (area ratio) of these confidence bands. From a viewpoint of the coverage probability, it is reasonable to use the uniform width- confidence bands (Gafarian bands) and trapezoidal confidence bands (Bowden-Graybill bands) for the finite interval of interest, rather than to use the Working-Hotelling bands. In particular, the use of these confidence bands can be emphasized when a length of the interval of interest is smaller than that of the observation interval. Moreover, from the examination about efficiency, it is strongly recommended to use the trapezoidal confidence bands when the interval of interest is near the relative ends of the observation interval.

View full abstract

We consider a problem of comparing several dose levels with a zero dose control and determining the highest dose (no-observed-adverse-effect level) at which there is no significant difference from the control in toxicological studies, when we can suppose that the toxic effect increases as the dose is increased. Under the normality assumption about the data, the authors have proposed the MDL test based on the MDL criterion formulating the problem as a model selection, and have shown the validity of the test. This paper extends the MDL test to a nonparametric case by replacing the data with their ranks, and gives approximate small-sample critical values for the test. Both probabilities of correctly and incorrectly determining the no-observed-adverse-effect level are compared between the proposed test and some competing tests by using a Monte Carlo simulation.

View full abstract

With the growing need for interdisciplinary and international availability, distribution and exchange of intellectual assets including information, knowledge, ideas, pieces of work, and tools in re-editable and redistributable organic forms, we need new media technologies that externalize scientific, technological, and/or cultural knowledge fragments in an organic way, and promote their advanced use, international distribution, reuse, and re-editing. These media may be called meme media since they carry what R. Dawkins called "memes". An accumulation of memes in a society forms a meme pool that functions like a gene pool. This paper reviews our R & D on meme media, and proposes their application framework for virtually materializing database records as interactive 3D objects. The 3D meme media architecture IntelligentBox provides the basis of this application framework. The materialization may use any interactive composite object defined in IntelligentBox as a virtual materialization template. You can associate the parameters of this composite object with arbitrarily selected attributes of each record.

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract