Plateau Phenomena of the Learning Dynamics and Stabilities of the Local Minima of the Error Function in Machine Learning

机器学习是构建能够像人脑一样学习外部世界提供的数据的系统的理论之一。机器学习的算法分为有监督学习、无监督学习和强化学习三类。在这项研究中，我们研究了监督学习和强化学习的动态，并提出了一些改进。(1)研究了隐马尔可夫模型参数空间的奇异结构与学习动力学轨迹之间的关系。(2)我们提出了强化学习算法，可以适应不断变化的环境和学习动力学研究。

Machine learning is one of the theories to construct the systems that can learn the data given from outside world like human brains. The algorithms of machine learning are divided into three categories such as supervised learning, unsupervised learning and reinforcement learning. In this research, we have investigated the dynamics of supervised learning and reinforcement learning and proposed some improvements. (1) We have investigated the relation between the singular structure of the parameter space of hidden Markov models and the trajectories of the learning dynamics. (2) We have proposed the reinforcement learning algorithm that can adapt to the changing environments and investigated the learning dynamics.