Reinforcement Learning Control with Time-Dependent Agent Dynamics
- Additional Document Info
- View All
Reinforcement learning-based control systems have the advantage of being modelfree. When used for control of dynamical systems, approximations of the time dynamics can be used to improve the policy development. This chapter proposes and develops a Q-Learning-based method for learning unknown information about the time scale properties of a dynamical system. Two candidate algorithm are proposed. The first determines optimal sample times for sampled data systems, and is demonstrated with a simulation of an inverted pendulum cart. The second algorithm learns a first-order approximation for a dynamical system, and uses the approximation to improve the control of a heterogenous multiagent system. This algorithm is demonstrated with a conceptual experiment involving such a system. © 2013 The Institute of Electrical and Electronics Engineers, Inc.
author list (cited authors)
Kirkpatrick, K., & Valasek, J.
Reinforcement Learning and Approximate Dynamic Programming for Feedback Control