Hierarchical Decomposition in Reinforcement Learning by Anders Jonsson
Hierarchical Decomposition in Reinforcement Learning
Author: Anders Jonsson
Title: Hierarchical Decomposition in Reinforcement Learning
ISBN10: 3836438615
ISBN13: 978-3836438612
Format: .PDF .EPUB .FB2
Pages:
Publisher: VDM Verlag Dr. Mueller e.K. (April 10, 2008)
Language: English
Size pdf: 1712 kb
Size epub: 1626 kb
Rating: 3.5 ✪
Votes: 321
Category: Computers & Technology
Subcategory: Computer Science
Reinforcement learning is an area of artificial intelligence that studies the ability of autonomous agents to improve their behavior in the absence of an informed instructor. Although reinforcement learning has achieved success in a wide range of applications, it becomes less consistent as the size of a task grows. This book attempts to improve the efficiency of reinforcement learning in realistic tasks by identifying a certain type of task structure. A task that displays this type of structure can be decomposed into a hierarchy of subtasks. Each subtask can be simplified using state abstraction so that it is much easier to solve than the original task. Reinforcement learning can be applied to produce solutions to the subtasks, and the solutions can be combined to achieve a solution to the original task. Experimental results indicate that hierarchical decomposition combined with state abstraction can significantly simplify the solution of realistic tasks. The book thus contributes to increasing the potential of reinforcement learning in realistic tasks. The book is directed towards researchers in Artificial Intelligence, but can also be used as a reference by professionals in Robotics and Autonomous Control Engineering.