PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Hierarchical Hybrid Reinforcement Learning Algorithms
Mohammad Ghavamzadeh
In: Workshop on Bridging the Gap between High-level Discrete Representations and Low-level Continuous Behaviors, Robotics: Science and Systems Conference (RSS-2009), 28 June 2009, Seattle, WA, USA.

Abstract

A possible approach to control high-dimensional problems, including those in robotics, is to design algorithms that can take advantage of the modularity of hierarchical reinforcement learning methods, the speed of value-function-based reinforcement learning algorithms, and the efficacy of policy gradient reinforcement learning methods in dealing with continuous state and action problems. In this work, we define a family of hierarchical reinforcement learning algorithms in which value-function-based reinforcement learning methods are used to control high-level subtasks, which usually involve smaller and more manageable state and finite action spaces, and policy gradient reinforcement learning controllers are used for low-level subtasks, which usually have continuous state and/or action spaces. We call this family of algorithms hierarchical hybrid reinforcement learning.

EPrint Type:Conference or Workshop Item (Poster)
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Learning/Statistics & Optimisation
ID Code:6096
Deposited By:Mohammad Ghavamzadeh
Deposited On:08 March 2010