PERFORMANCE COMPARISON OF REINFORCEMENT LEARNING ALGORITHMS ON CART-POLE CONTROL PROBLEM

Özgür H. E., Sarıgeçili M. İ.

CISET - 2nd Cilicia International Symposium on Engineering and Technology, Mersin, Turkey, 10 - 12 October 2019, pp.262-265

  • Publication Type: Conference Paper / Full Text
  • City: Mersin
  • Country: Turkey
  • Page Numbers: pp.262-265
  • Çukurova University Affiliated: Yes

Abstract

Machine learning-based control is an emerging and promising area in control applications. Reinforcement Learning is an attractive part of machine learning. In this study, five different Reinforcement Learning algorithms (i.e. Deep-Q Networks, Trusted Region Policy Optimization, Proximal Policy Optimization, Asynchronous Advantage Actor-Critic and Actor-Critic using Kronecker-Factored Trust Region) and their performances at three different levels of time step (i.e. 105, 106 and 107) are presented on the inverted pendulum on cart since it is a basic problem studied widely in control.
Machine learning-based control is an emerging and promising area in control applications. Reinforcement Learning is an attractive part of machine learning. In this study, five different Reinforcement Learning algorithms (i.e. Deep-Q Networks, Trusted Region Policy Optimization, Proximal Policy Optimization, Asynchronous Advantage Actor-Critic and Actor-Critic using Kronecker-Factored Trust Region) and their performances at three different levels of time step (i.e. 105, 106 and 107) are presented on the inverted pendulum on cart since it is a basic problem studied widely in control.