CMA-TWEANN: Efficient optimization of neural networks via self-adaptation and seamless augmentation

Hirotaka Moriguchi, Shinichi Honiden

Research output: Chapter in Book/Report/Conference proceedingConference contribution

11 Citations (Scopus)

Abstract

Neuroevolutionary algorithms are successful methods for optimizing neural networks, especially for learning a neural policy (controller) in reinforcement learning tasks. Their significant advantage over gradient-based algorithms is the capability to search network topology as well as connection weights. However, state-of-the-art topology evolving methods are known to be inefficient compared to weight evolving methods with an appropriately hand-tuned topology. This paper introduces a novel efficient algorithm called CMA-TWEANN for evolving both topology and weights. Its high efficiency is achieved by introducing efficient topological mutation operators and integrating a state-of-the-art function optimization algorithm for weight optimization. Experiments on benchmark reinforcement learning tasks demonstrate that CMA-TWEANN solves tasks significantly faster than existing topology evolving methods. Furthermore, it outperforms weight evolving techniques even when they are equipped with a hand-tuned topology. Additional experiments reveal how and why CMA-TWEANN is the best performing weight evolving method.

Original languageEnglish
Title of host publicationGECCO'12 - Proceedings of the 14th International Conference on Genetic and Evolutionary Computation
Pages903-910
Number of pages8
DOIs
Publication statusPublished - 2012 Aug 13
Externally publishedYes
Event14th International Conference on Genetic and Evolutionary Computation, GECCO'12 - Philadelphia, PA, United States
Duration: 2012 Jul 72012 Jul 11

Publication series

NameGECCO'12 - Proceedings of the 14th International Conference on Genetic and Evolutionary Computation

Other

Other14th International Conference on Genetic and Evolutionary Computation, GECCO'12
CountryUnited States
CityPhiladelphia, PA
Period12/7/712/7/11

Keywords

  • neuroevolution
  • reinforcement leanring
  • tweann

ASJC Scopus subject areas

  • Computational Theory and Mathematics
  • Applied Mathematics

Fingerprint Dive into the research topics of 'CMA-TWEANN: Efficient optimization of neural networks via self-adaptation and seamless augmentation'. Together they form a unique fingerprint.

Cite this