Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

Control of exploitation-exploration meta-parameter in reinforcement learning.

Shin Ishii1, Wako Yoshida, Junichiro Yoshimoto

  • 1Nara Institute of Science and Technology, Ikoma, Japan. ishii@is.aist-nara.ac.jp

Neural Networks : the Official Journal of the International Neural Network Society
|October 10, 2002
PubMed
Summary
This summary is machine-generated.

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Enhancing experimental design through Bayes factor design analysis: insights from multi-armed bandit tasks.

Wellcome open research·2026
Same author

Template-Based Label Propagation for Mouse Brain MRI Skull Stripping.

Neuroinformatics·2026
Same author

Machine learning-based mortality prediction for pediatric fulminant myocarditis using cytokine profiles.

Scientific reports·2026
Same author

Inferring relationships among major psychiatric disorders in a resting-state functional connectivity-informed embedding space.

NPJ systems biology and applications·2026
Same author

Insula Structure Is Linked to Autonomic Cardiac Dysregulation in Depression.

Biological psychiatry·2026
Same author

CRTC1 knockdown in the marmoset visual cortex induces neuronal IEG overexpression, HFOs, and neurodegeneration.

Neuroscience research·2026

This study introduces a novel reinforcement learning (RL) method to balance exploitation and exploration. The approach adapts to environmental changes, demonstrating effective control in maze tasks and offering insights into brain mechanisms.

Area of Science:

  • Artificial Intelligence
  • Computational Neuroscience
  • Machine Learning

Background:

  • The exploitation-exploitation dilemma is a fundamental challenge in reinforcement learning (RL).
  • Existing RL methods often struggle to dynamically adapt to changing environments.
  • Understanding the neural basis of decision-making, particularly the balance between exploration and exploitation, is crucial.

Purpose of the Study:

  • To develop a novel model-based reinforcement learning method for dynamically controlling the exploitation-exploration balance.
  • To investigate how environmental changes influence optimal action selection strategies.
  • To explore potential neural implementations of this balance control mechanism.

Main Methods:

  • Utilized model-based reinforcement learning with Bayes inference and a forgetting effect to estimate environmental state-transition probabilities.

Related Experiment Videos

  • Developed a balance parameter to control action selection randomness, adapting to variations in action outcomes and perceived environmental shifts.
  • Tested the method on maze tasks to evaluate its adaptive control capabilities.

    • Main Results:

    • The proposed method successfully achieved adaptive control in maze tasks.
    • The system demonstrated an ability to adjust the exploitation-exploration balance in response to environmental changes.
    • The findings suggest a viable computational framework for balancing exploration and exploitation.

    Conclusions:

    • The novel RL method effectively manages the exploitation-exploration trade-off by dynamically adjusting a balance parameter.
    • The approach shows promise for applications requiring adaptive decision-making in changing environments.
    • The study provides a potential computational model for understanding the brain's control of selective attention and exploration-exploitation balance.