Improving Behaviour by Modelling Irrelevant Environmental Features
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 a model based algorithm that has only recently been created.
It uses a compact latent state to store information about the current state
of the environment to gain higher rewards. We observed that DreamerV2
has a problem in more complex environments because it is unable to encode
all the features of this more complex environment properly. This research
introduces an alteration of DreamerV2 that can work better in these certain
environments. We created a CNN that gets multiple images, with these
images the CNN tries to predict the irrelevant features of the environment.
By adding this CNN to the DreamerV2 framework we expect the model
to get higher rewards when the complexity of the environment increases.
We expect that, because the CNN predicts all irrelevant features, only the
relevant features need to be processed by the auto-encoder of DreamerV2,
resulting in a higher reward when playing a game. We have seen that our
model does exactly this. Our added model predicts some of the irrelevant
features, resulting in a significantly higher reward for our model,compared to
the original DreamerV2, when trained on a game with an added distractor.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen