Separating relevant from irrelevant environmental factors in model-based reinforcement learning
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, a state of the art model-based reinforcement learning algorithm,
has shown that it is a viable competitor to model-free reinforcement
learning algorithms on the Atari 55 benchmark. It does so by learning a
world model and training an agent using only the world model. DreamerV2
does fail when extra distractions are added to the environments. To combat
this issue we investigated whether separating relevant from irrelevant
environmental factors improves performance of DreamerV2. Adding a convolutional
neural network to facilitate this separation and make the encoder
focus on only the relevant parts. It turns out that using a 2D CNN with
instance normalization works best for separating information compared to
a 3D CNN and different normalization techniques such as batch and act
normalization.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen