Browsing by Author "Gao, Yuan"

Controlling a complicated mechanical system to perform a certain task, for example, making robot to dance, is a traditional problem studied in the area of control theory. However, evidence shows that incorporating machine learning techniques in robotics can enable researchers to get rid of tedious engineering works of adjusting environmental parameters. Many researchers like Jan Peters, Sethu Vijayakumar, Stefan Schaal, Andrew Ng and Sebastian Thrun are the early explorers in this field. Based on the Partial Observable Markov Decision Process (POMDP) reinforcement learning, they contributed theory and practical implementation of several benchmarks in this field. Recently, one sub-field of machine learning called deep learning gained a lot of attention as a method attempting to model high-level abstractions by using model architectures composed of multiple non-linear layers (for example [Krizhevsky2012]). Several architectures of deep learning networks like deep belief network [Hinton2006], deep Boltzman machine [Salakhutdinov2009], convolutional neural network [Krizhevsky2012] and deep de-noising auto-encoder [Vincent2010] have shown their advantages in specific areas. The main contribution of deep learning is more related to perception which deals with problems like Sensor Fusion [OConnor2013], Nature Language Processing(NLP)[Cho2014] and Object Recognition [Lenz2013][Hoffman2014]. Although considered briefly in Jürgen Schmidhuber's team[Mayer2006], the other area of robotics, namely control, remains more-or-less unexplored in the realm of deep learning. The main focus of this thesis is to introduce general learning methods for robot control problem with an exploration on deep learning method. As a consequence, this thesis tries to describe the transitional learning methods as well as the emerging deep learning methods including new findings in the investigation.