Hiroyuki Shioya

Empirical analysis of an on-line adaptive system using a mixture of Bayesian networks

An on-line reinforcement learning system that adapts to environmental changes using a mixture of Bayesian networks is described. Building intelligent systems able to adapt to dynamic environments is important for deploying real-world applications. Machine learning approaches, such as those using reinforcement learning methods and stochastic models, have been used to acquire behavior appropriate to environments characterized by uncertainty. However, efficient hybrid architectures based on these approaches have not yet been developed. The results of several experiments demonstrated that an agent using the proposed system can flexibly adapt to various kinds of environmental changes.

Low voltage electron diffractive imaging of atomic structure in single-wall carbon nanotubes

The demand for atomic-scale analysis without serious damage to the specimen has been increasing due to the spread of applications with light-element three-dimensional (3D) materials. Low voltage electron diffractive imaging has the potential possibility to clarify the atomic-scale structure of 3D materials without causing serious damage to specimens. We demonstrate low-voltage (30 kV) electron diffractive imaging of single-wall carbon nanotube at a resolution of 0.12 nm. In the reconstructed pattern, the intensity difference between single carbon atom and two overlapping atoms can be clearly distinguished. The present method can generally be applied to other materials including biologically important ones.

Spherical shell structure of distribution of images reconstructed by diffractive imaging.

Image reconstruction from Fourier intensity through phase retrieval was investigated when the intensity was contaminated with Poisson noise. Although different initial conditions and/or the instability of the iterative phase retrieval process led to different reconstructed images, we found that the distribution of the resulting images in both the object and Fourier spaces formed spherical shell structures. Averaging of the images over the distribution corresponds to the position of the image at the sphere center.

Unsupervised Weight Parameter Estimation Method for Ensemble Learning

When there are multiple trained predictors, one may want to integrate them into one predictor. However, this is challenging if the performances of the trained predictors are unknown and labeled data for evaluating their performances are not given. In this paper, a method is described that uses unlabeled data to estimate the weight parameters needed to build an ensemble predictor integrating multiple trained component predictors. It is readily derived from a mathematical model of ensemble learning based on a generalized mixture of probability density functions and corresponding information divergence measures. Numerical experiments demonstrated that the performance of our method is much better than that of simple average-based ensemble learning, even when the assumption placed on the performances of the component predictors does not hold exactly.

Design of an Unsupervised Weight Parameter Estimation Method in Ensemble Learning

A learning method using an integration of multiple component predictors as an ultimate predictor is generically referred to as ensemble learning. The present paper proposes a weight parameter estimation method for ensemble learning under the constraint that we do not have any information of the desirable (true) output. The proposed method is naturally derived from a mathematical model of ensemble learning, which is based on an exponential mixture type probabilistic model and Kullback divergence. The proposed method provides a legitimate strategy for weight parameter estimation under the abovementioned constraint if it is assumed that the accuracy of all multiple predictors are the same. We verify the effectiveness of the proposed method through numerical experiments.

Phase retrieval based on an Evolutionary Multicriterion Optimisation method

Phase problems arise from lost phase information in measurement of diffraction waves. The missing phase should be retrieved to reconstruct an object image from the diffraction pattern. This paper proposes a hybrid type approach, Evolutionary-based GS (E-GS), based on the Gerchberg — Saxton algorithm (GS algorithm) and Evolutionary Multicriterion Optimisation (EMO). There are three main aims of E-GS: (1) to reduce the dependence on initial conditions, (2) to obtain some candidate solutions with various features in one trial and (3) to achieve algorithmic parallelism. In E-GS, the phase retrieval problem is formulated as a two-objective optimisation problem, and the EMO and GS algorithm are used as the framework of multiobjective optimisation and local search, respectively. E-GS deals directly with phase as an optimisation parameter and embeds original genetic operations based on frequency characteristics. In this paper, the characteristics and effectiveness of the proposed approach are discussed by comparison of the performance with that of the GS algorithm. Through numerical examples, it was demonstrated that E-GS could derive good results and the difference of search transition between GS algorithm and E-GS was clarified.

Invited Paper: An Information-Theoretic Approach to Phase Retrieval

Phase problems arise from the lost Fourier phase in measuring the diffraction waves. Re-constructing the phase information using the diffraction pattern of a target object yields the target image, and it is called phase retrieval. This paper introduces an information-theoretic approach to phase retrieval based on information measures, and a refined derivation of the generalized phase retrieval algorithm based on the density power divergence is presented with a simple numerical example using the Poisson-noise-contaminated Fourier intensity.

A Policy-Improving System with a Mixture of Bayesian Networks Adapting Agents to Continuously Changing Environments

A variety of adaptive learning systems which adapt themselves to complicated environments has been studied and developed in the broad field of AI researches. For example, many reinforcement learning (RL) methods have been proposed to adapt agents to the environments. At the same time, Bayesian network (BN), one of the stochastic models, has attracted increasing attention due to its noise robustness, reasoning power, etc. We have proposed a system improving RL agents policies with a mixture model of RNs, and have evaluated the adapting performance of our system. Each structure of BN can be regarded as a stochastic knowledge representation in the policy acquired through RL. It has been confirmed that the agent with our system could improve their policies by the information derived from the mixture, and then could adequately adapt to dynamically-switched environments. In this research, we propose a method to appropriately normalize mixing parameters of the mixture for the use in common adaptive learning systems, and evaluate the fundamental performance of our system in continuously-changing environment