Towards Out-of-distribution Problem for Reinforcement Learning

Deep neural models have achieved impressive success and progress in the last decade. However, high-quality models require a large amount of data, parameters as well as computation power. This originates from the curse of dimensionality and poor out-of-distribution generalization of current probabilistic models. Current machine learning models requires data points to be independently identically distributed which is often not satisfied in real-world applications. This mismatch damages the direct application of classic learning models on out-of-distribution data. In this dissertation, we propose to explore this issue from three perspectives. First, we explore the impact of distribution perturbation under the adversarial attack, which validates the sensitivity of deep learning models under even small distribution shifts. To increase the robustness of our system, we propose a detection model in the recommendation system scenario. The second problem we investigate is the domain adaptation. Specifically, we study how to learn good representations to map samples from one domain to another domain in the image transfer setting. Finally, we probe into the domain generalization setting where a model aims to achieve better performance under multiple domains. We study the meta-learning models to learn directly from multi-task settings to explore a way to learn representation under a large distribution shift

Grantor dc:publisher

UNSW, Sydney

Year dc:date

2023

Author dc:creator

• Cao, Yuanjiang

• out-of-distribution generalization
• reinforcemenet learning
• anzsrc-for: 461105 Reinforcement learning

Statement dc:rights

• open access
• CC BY 4.0
• free_to_read

Licence

• https://purl.org/coar/access_right/c_abf2
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

en

Identifier

https://doi.org/10.26190/unsworks/24998

OAI identifier oai:identifier

oai:unsworks.library.unsw.edu.au:1959.4/101291