Evolutionary reinforcement learning for simulated locomotion of a robot with a two-link arm.

Y. Kassahun, G. Sommer
IAS-9
Proceedings of the 9th Conference on Intelligent Autonomous Systems (IAS-9)
Pages 263-271
March, 2006

Abstract

In this paper we present a neural controller design for robots using an evolutionary reinforcement learning system that is suitable for learning through interaction. The method starts with networks of minimal structures determined by the domain expert and increases their complexity along the evolution path. It uses a nature inspired meta-level evolutionary process where new structures are explored at larger time-scale and existing structures are exploited at smaller time-scale. The method introduces an efficient and compact genetic encoding of neural networks onto a linear genome that enables one to evaluate the network without decoding it. We demonstrate the method by designing a neural controller for a robot with a two-link arm that enables it to move forward as fast as possible. We first give an introduction to the evolutionary method and then describe the experiment and results obtained.

Keywords

Neural networks, Reinforcement learning, Evolutionary algorithms

Full Paper

Portable document format file PDF

Bibtex entry

@InProceedings{kassahun06ias,
  author =       {Y. Kassahun and G. Sommer},
  title =        {Evolutionary reinforcement learning for simulated locomotion
                  of a robot with a two-link arm.},
  booktitle =    {Proceedings of the 9th Conference on Intelligent Autonomous
                  Systems (IAS-9)},
  year =         {2006},
  month =        {March},
  editor =       {T. Arai and R. Pfeifer and T. Balch and H. Yokoi},
  publisher =    {IOS Press},
  pages =        {263--271}
}