強化学習における次元削減

本研究的主题是强化学习中的降维问题。到目前为止，我们的成就是三个出版物。第一篇论文发表在《神经网络》杂志，第84卷，第1-16页，2016年12月。在这项工作中，我们提出了一种基于模型的降维强化学习方法。我们与京都ATR研究所的脑机接口部门合作，并在一个全尺寸的类人机器人上对所提出的方法进行了实验评估。第二篇发表在2017年2月AAAI人工智能会议论文集上。在这项工作中，我们提出了基于模型的降维上下文强化学习方法。我们与德国达姆施塔特工业大学的自主智能系统实验室合作，并使用模拟的机械臂对所提出的方法进行了实验评估。第三篇论文被《神经计算》杂志接受，将于2017年3月21日出版。在这项工作中，我们给出了有监督学习环境下基于二次互信息的降维方法的初步结果。

The topic of this research is dimension reduction in reinforcement learning. Our achievements so far are three publications.1. The first publication appeared in Neural Networks journal, vol. 84, pages 1-16, December 2016. In this work we developed a model-based reinforcement learning method with dimension reduction. We collaborate with the Brain-Computer Interface department at ATR institute in Kyoto and experimentally evaluate the proposed method on a full-size humanoid robot.2. The second publication appeared in the proceedings of AAAI conference on Artificial Intelligence, February 2017. In this work we develop model-based contextual reinforcement learning method with dimension reduction. We collaborate with the Autonomous Intelligent Systems laboratory at Technical University of Darmstadt in Germany and experimentally evaluate the proposed method using a simulated robot arm.3. The third publication is accepted for publication by the Neural Computation journal on 21st March 2017. In this work, we presented preliminary results on dimension reduction method based on quadratic mutual information in the supervised learning setting.