In this paper, we propose a method of designing input-output history feedback controllers for unknown linear discrete-time systems. Many conventional reinforcement-learning based controls such as policy iteration are state-feedback. We extend the policy iteration by incorporating a method to statically estimate state variables from a history of finite-time input-output data. The convergence of the policy to model-based optimal solution has been theoretically guaranteed. Moreover, the proposed method is one-shot learning, i.e., the optimal controller can be obtained by using initial experiment data only. The effectiveness of the proposed method is shown through a numerical simulation through an oscillator network.
In this paper, we applied topological data analysis (TDA) to the segmentation of grayscale images. TDA is a method to extract topological features such as hollows from data. It is expected that we can segment images because we can regard objects in images as hollows. In this paper, we first confirmed that the TDA based method was effective in the segmentation of halftone images by random dithering. Then, we compared the proposed method and the combination of Otsu’s method and TDA. Finally, we evaluated the performance of our method using standard images and CT-images.
Mobile robot localization in outdoor is difficult because of environmental change such as sunlight intensity, moving obstacles layout, etc. In this article, questionnaire study on human ability of environmental recognition was carried out and its results were applied to mobile robot localization. The result of questionnaire shows that a human relies on globally visual information such as landscape rather than information of landmarks such as signboards. Based on the result, a view-based localization method using GIST features was proposed. GIST is known to represent well the overview of outdoor scene. A capture scene image is compared all registered template images in terms of GIST, e.g., the template image, of which L2 norm of GIST between input image is minimum, is successively searched. Then, the robot is localized so as that it exists in the neighborhood of known place where the template image was captured. Experimental result in the course of Tsukuba Challenge 2015 showed the good potential of proposed method to localize a mobile robot.