Federated Offline Reinforcement Learning for Autonomous Systems

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Offline reinforcement learning (RL) provides a safe learning method that can be applied to real-world applications through a data-driven learning process. In general, this process attributes to learning on large datasets, which is similar to conventional supervised learning techniques. In this paper, we propose a federated offline RL framework where policy parameters learned locally on separate datasets of RL agents are aggregated in communication rounds to achieve a high-performance global policy. The framework can solve the data insufficiency issue by performing offline RL among multiple agents without data sharing and integration. We show that using the framework, agents achieve a stable-performance control policy in a 2D navigation environment, even when each has only 1% of the required training trajectory amount (e.g., 10 trajectories).

Original languageEnglish
Title of host publicationComputer and Communication Engineering - 2nd International Conference, CCCE 2022, Revised Selected Papers
EditorsFilippo Neri, Ke-Lin Du, Vijaya Kumar Varadarajan, San-Blas Angel-Antonio, Zhiyu Jiang
PublisherSpringer Science and Business Media Deutschland GmbH
Pages107-117
Number of pages11
ISBN (Print)9783031174216
DOIs
StatePublished - 2022
Event2nd International Conference on Computer and Communication Engineering, CCCE 2022 - Virtual, Online
Duration: 11 Mar 202213 Mar 2022

Publication series

NameCommunications in Computer and Information Science
Volume1630 CCIS
ISSN (Print)1865-0929
ISSN (Electronic)1865-0937

Conference

Conference2nd International Conference on Computer and Communication Engineering, CCCE 2022
CityVirtual, Online
Period11/03/2213/03/22

Keywords

  • Autonomous systems
  • Federated learning
  • Offline reinforcement learning

Fingerprint

Dive into the research topics of 'Federated Offline Reinforcement Learning for Autonomous Systems'. Together they form a unique fingerprint.

Cite this