Deep vs. Deep Bayesian: Faster Reinforcement Learning on a Multi-robot Competitive Experiment

Jingyi Huang, Fabio Giardina, Andre Rosendo

2021

Abstract

Deep Learning experiments commonly require hundreds of trials to properly train neural networks, often labeled as Big Data, while Bayesian learning leverages scarce data points to infer next iterations, also known as Micro Data. Deep Bayesian Learning combines the complexity from multi-layered neural networks to probabilistic inferences, and it allows a robot to learn good policies within few trials in the real world. In here we propose, for the first time, an application of Deep Bayesian Reinforcement Learning (RL) on a real-world multi-robot confrontation game, and compare the algorithm with a model-free Deep RL algorithm, Deep Q-Learning. Our experiments show that DBRL significantly outperforms DRL in learning efficiency and scalability. The results of this work point to the advantages of Deep Bayesian approaches in bypassing the Reality Gap and sim-to-real implementations, as the time taken for real-world learning can quickly outperform data-intensive Deep alternatives.

Paper Citation

in Harvard Style

Huang J., Giardina F. and Rosendo A. (2021). Deep vs. Deep Bayesian: Faster Reinforcement Learning on a Multi-robot Competitive Experiment. In Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics - Volume 1: ICINCO, ISBN 978-989-758-522-7, pages 501-506. DOI: 10.5220/0010601905010506

in Bibtex Style

@conference{icinco21,
author={Jingyi Huang and Fabio Giardina and Andre Rosendo},
title={Deep vs. Deep Bayesian: Faster Reinforcement Learning on a Multi-robot Competitive Experiment},
booktitle={Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics - Volume 1: ICINCO,},
year={2021},
pages={501-506},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0010601905010506},
isbn={978-989-758-522-7},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics - Volume 1: ICINCO,
TI - Deep vs. Deep Bayesian: Faster Reinforcement Learning on a Multi-robot Competitive Experiment
SN - 978-989-758-522-7
AU - Huang J.
AU - Giardina F.
AU - Rosendo A.
PY - 2021
SP - 501
EP - 506
DO - 10.5220/0010601905010506