University of Twente Student Theses

Login

Data augmentation for regularizing learned world models in reinforcement learning

Bree, R.J. van (2021) Data augmentation for regularizing learned world models in reinforcement learning.

[img] PDF
631kB
Abstract:This work has investigated the effect of data augmentation on learning environment models for data-efficient reinforcement learning. Data augmentation has received relatively little attention in reinforcement learning, compared to its common occurrence in supervised learning. However, the invariance assumptions that allow the augmentation of images have long been recognized to be usable for the simplification of reinforcement learning problems. Recent successes in applying data augmentation to model-free reinforcement learning algorithms raised the question whether data augmentation could be effective enough to enforce these assumptions for learned environment models, for an increase in data efficiency. This has been numerically investigated using an algorithm that is currently among the state of the art for learning environment models, called the Recurrent State Space Model (RSSM). The use of data augmentation was found to have no significant effects on the consistency of the RSSM when faced with state-action sequences that have an equivalent reward expectation.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:http://purl.utwente.nl/essays/85596
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page