Year: 2015
Type: dissertation
Abstract: Reinforcement Learning (RL) is a framework to deal with decision-making problems with the goal of finding optimal control policies. Robotics decision-making problems are of importance and provide interesting applications for reinforcement learning; however, since real-world robots have many degrees of freedom, high-dimensionality arises when applying reinforcement learning algorithms. Deep Neural Networks (DNNs) have the potential to reduce the high-dimensional state spaces without losing information. A low-dimensional representation is generated by training a multilayer neural network with a small central layer to reconstruct high-dimensional input vectors. Back- propagation algorithms can be used to tune the weightings in this neural network with the help of the gradient descent method. The objective of this project is to develop approaches of applying deep neural networks with reinforcement learning, hereby, efficiently generating dynamic policies in robot motion-control ap- plications. More specifically, the objective of this project is to overcome the high-dimensionality challenge in reinforcement learning with the help of deep neural networks, a stacked auto-encoder, and using various types of sensory inputs like angles and angular velocities in the manipulator. Experimentally, a simple segway system, which consists of a link and a wheel, is used to test per- formance. The main approach to improve the learning performance is to develop a low-dimensional representation replacing the original inputs that describe the current states of the segway system in the learning process and using a stacked auto-encoder instead.
Author: Xiaoliang Jia
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1904233931