Skip to main navigation Skip to search Skip to main content

Selective Transition Collection in Experience Replay

  • Beihang University

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

Abstract

Experience replay method is often used in off-policy reinforcement learning. As the training progresses, the distribution of the collected transitions becomes more and more concentrated, and this will lead to catastrophic forgetting and a low rate of convergence. In this paper, we present selective transition collection algorithm which is a new design to address the concentrated distribution by selectively collection the transitions. We propose a method to estimate the similarity between transitions, and a probability function to reduce the chance of transitions with high similarity to the experience memory being collected. We test our method on familiar reinforcement learning tasks and the experimental results demonstrate that selective transition collection can not only speed up the learning but also prevent catastrophic forgetting effectively.

Original languageEnglish
Title of host publicationCognitive Systems and Signal Processing - 5th International Conference, ICCSIP 2020, Revised Selected Papers
EditorsFuchun Sun, Huaping Liu, Bin Fang
PublisherSpringer Science and Business Media Deutschland GmbH
Pages317-324
Number of pages8
ISBN (Print)9789811623356
DOIs
StatePublished - 2021
Event5th International Conference on Cognitive Systems and Signal Processing, ICCSIP 2020 - Zhuhai, China
Duration: 25 Dec 202027 Dec 2020

Publication series

NameCommunications in Computer and Information Science
Volume1397 CCIS
ISSN (Print)1865-0929
ISSN (Electronic)1865-0937

Conference

Conference5th International Conference on Cognitive Systems and Signal Processing, ICCSIP 2020
Country/TerritoryChina
CityZhuhai
Period25/12/2027/12/20

Keywords

  • Experience replay
  • Reinforcement learning
  • Selective transition collection

Fingerprint

Dive into the research topics of 'Selective Transition Collection in Experience Replay'. Together they form a unique fingerprint.

Cite this