Improving DreamerV2 by increasing its resistance to distractions
Keywords
Loading...
Authors
Issue Date
2022-06-20
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
DreamerV2 is able to achieve human-level performance on the Atari benchmark.
It accomplishes this by learning an accurate model of the world in an
efficient latent space. However, the model encounters a problem when the
environments become more complex. The model gets distracted by many
environmental features and complex backgrounds which results in it being
unable train the agent accurately. This paper introduces an alteration on
DreamerV2 that aims to solve this problem. By adding a Convolutional
Neural Network (CNN) that combines previous observations of the world
into a single picture, it is able to extract some of the irrelevant features
of the world. By supplying this information to DreamerV2, the model can
focus on learning the relevant aspects of the environment that are needed
to obtain the rewards. This model is able to obtain comparable results to
DreamerV2 on a relatively complex game and it significantly improves on
DreamerV2 when a large distractor is added to the game.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen