Using probabilistic modelling to improve performance of reinforcement learning agents for tra c ow optimization problems in partially observable environments
Keywords
Loading...
Authors
Issue Date
2021-06-18
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Increasing tra c volumes throughout the world call for controlling infrastructure
in such a way that optimal tra c
ow is achieved. Several
methods suggest reinforcement learning based tra c light programs, which
seem to give a major tra c
ow improvement in controlled scenarios. However,
in real world scenarios, sensor data is much more scarce, it may not
be possible to collect data which is as complete as the data most of the
suggested methods receive. This calls for a new approach which aims to
alleviate the problems raised by missing or inaccurate sensor data. In this
thesis, a variational auto encoder using graph neural networks is proposed,
which will be used to reconstruct missing sensor data. The reconstructed
data can subsequently be used as input for existing reinforcement learning
methods, which has shown to be an improvement in some cases. The model
has shown to improve performance of one existing method, however, more
research has to be conducted in order to draw nal conclusions.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen