Approximation of Agent Dynamics Using Reinforcement Learning
Conference Paper
Overview
Identity
Other
View All
Overview
abstract
Reinforcement Learning for control of dynamical systems is popular due to the ability to learn control policies without requiring a model of the system being controlled. It can be difficult to learn ideal control policies because it is common to abstract out or ignore completely the dynamics of the agents in the system. In this paper, Reinforcement Learning-based algorithms are developed for learning agents' time dependent dynamics while also learning to control them. Three algorithms are introduced. Sampled-Data Q-learning is an algorithm that learns the optimal sample time for controlling an agent without a prior model. First-Order Dynamics Learning is an algorithm that determines the proper time constants for agents known to have first-order dynamics, while Second-Order Dynamics Learning is an algorithm for learning natural frequencies and damping ratios of second-order systems. The algorithms are demonstrated with numerical simulation. Results presented in this paper show that the algorithms are able to determine information about the system dynamics without resorting to traditional system identification. 2013 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
name of conference
51st AIAA Aerospace Sciences Meeting including the New Horizons Forum and Aerospace Exposition
